ArticlePDF Available

Abstract and Figures

We introduce HammerDrive, a novel architecture for task-aware visual attention prediction in driving. The proposed architecture is learnable from data and can reliably infer the current focus of attention of the driver in real-time, while only requiring limited and easy-to-access telemetry data from the vehicle. We build the proposed architecture on two core concepts: 1) driving can be modeled as a collection of sub-tasks (maneuvers), and 2) each sub-task affects the way a driver allocates visual attention resources, i.e., their eye gaze fixation. HammerDrive comprises two networks: a hierarchical monitoring network of forward-inverse model pairs for sub-task recognition and an ensemble network of task-dependent convolutional neural network modules for visual attention modeling. We assess the ability of HammerDrive to infer driver visual attention on data we collected from 20 experienced drivers in a virtual reality-based driving simulator experiment. We evaluate the accuracy of our monitoring network for sub-task recognition and show that it is an effective and lightweight network for reliable real-time tracking of driving maneuvers with above 90% accuracy. Our results show that HammerDrive outperforms a comparable state-of-the-art deep learning model for visual attention prediction on numerous metrics with ∼13% improvement for both Kullback-Leibler divergence and similarity, and demonstrate that task-awareness is beneficial for driver visual attention prediction.
Content may be subject to copyright.
Preprint version; final version available at https://ieeexplore.ieee.org/
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1
HammerDrive: A Task-Aware Driving
Visual Attention Model
Pierluigi Vito Amadori, Member, IEEE, Tobias Fischer, Member, IEEE, Yiannis Demiris, Senior Member, IEEE
Abstract—We introduce HammerDrive, a novel architecture
for task-aware visual attention prediction in driving. The pro-
posed architecture is learnable from data and can reliably infer
the current focus of attention of the driver in real-time, while
only requiring limited and easy-to-access telemetry data from the
vehicle. We build the proposed architecture on two core concepts:
1) driving can be modeled as a collection of sub-tasks (maneuvers),
and 2) each sub-task affects the way a driver allocates visual
attention resources, i.e., their eye gaze fixation. HammerDrive
comprises two networks: a hierarchical monitoring network
of forward-inverse model pairs for sub-task recognition and
an ensemble network of task-dependent convolutional neural
network modules for visual attention modeling. We assess the
ability of HammerDrive to infer driver visual attention on data
we collected from 20 experienced drivers in a virtual reality-based
driving simulator experiment. We evaluate the accuracy of our
monitoring network for sub-task recognition and show that it
is an effective and light-weight network for reliable real-time
tracking of driving maneuvers with above 90% accuracy. Our
results show that HammerDrive outperforms a comparable state-
of-the-art deep learning model for visual attention prediction on
numerous metrics with 13% improvement for both Kullback-
Leibler divergence and similarity, and demonstrate that task-
awareness is beneficial for driver visual attention prediction.
Index Terms—Advanced Driver-Assistance Systems, Visual
Attention, Task Recognition, Simulated Driving, HAMMER.
I. INTRODUCTION
D
ISTRACTION, or misplaced attention, is regarded as the
leading cause of vehicle accidents [
1
]. Several studies
have confirmed this and have found distinct connections
between driving accidents to some form of distraction [
2
],
[
3
]. While advances in computing have fueled increasingly
complex and intelligent systems for active safety assistance in
driving, active monitoring still represents a challenge for the
deployment of Advanced Driver-Assistance Systems (ADAS).
Active monitoring systems are required to timely and reliably
evaluate both the driver’s actual focus of attention and the ideal
focus of attention for the driving task [4], [5].
In this paper, we introduce a model for visual attention
prediction in driving. In line with the literature [
5
], [
6
], we
formalize the problem as human visual attention modeling,
or eye fixation prediction, which has been an active research
topic in computer vision, robotics and neuroscience for many
years [
7
]. Human visual attention modeling can be described as
the task of inferring the focus of attention of a human observer
when looking at images or videos. Led by the pioneering work
P.V. Amadori, T. Fischer and Y. Demiris are with the Personal Robotics
Lab, Dept. of Electrical & Electronic Engineering, Imperial College London.
Manuscript received April 28, 2020; revised October 01, 2020, November
26, 2020; accepted January 15, 2021.
by Itti et al. [
8
], many studies have focused on designing
computational vision attention models that can predict human
eye fixations in static image observation [9], [10].
The advent of deep neural networks, together with large-
scale publicly available datasets and benchmarks, have further
improved static visual attention models, up to the point where
it is not possible to differentiate model predictions from
human fixation maps [
9
], [
11
]. However, visual attention
models for static image viewing cannot address nor leverage
the known correlation between human fixation patterns and
time [
12
]. Static-scene viewing models assume the observer to
have multiple seconds of observation over a single image,
while dynamic-viewing, as in driving, is characterized by
significantly shorter times per frame [
13
]. Behavioral studies
have also shown that motion is a key component in human
attention [
14
]. The ability to predict human eye fixations in
dynamic environments has several real-world applications,
spanning from attention-based video compressing [
15
] to
human-robot-interactions [16], [17].
In this paper, we focus on visual attention prediction in
highway driving, where it offers significant applications for
ADAS [
18
], such as blind spot control, distraction detection
or lane-change assistance [
19
], [
20
]. Recently, ADAS have
experienced an increasing interest in research as they aim to
improve driving safety and comfort. While active safety fea-
tures, e.g., collision avoidance and lane change assistance [
21
],
[
22
], have achieved notable success, many challenges still exist
for driver monitoring-based assistance [
4
], [
23
]. Among these,
the major source of complexity is the need to be able to predict
future intentions of drivers and their attention allocation in
a continuous and reliable manner [
4
]. Furthermore, ADAS
have additional computational constraints, as they need to
be deployed on platforms with limited resources, such as
embedded systems [24].
In this context, we introduce a novel architecture, namely
HammerDrive, for visual attention prediction in simulated
driving. The proposed architecture exploits real-time ma-
neuver recognition for task-aware visual attention modeling,
as overviewed in Fig. 1. HammerDrive uses easy-to-access
telemetry data from the vehicle, i.e., location, steering angle,
speed and throttle, to perform reliable and real-time prediction
of visual focus of the driver. Driving maneuver recognition
is performed by a Hierarchical, Attentive, Multiple Models
for Execution and Recognition (HAMMER) [
25
] network,
which is a general framework for recognizing and executing
actions by selecting modules depending on their prediction
error. Visual attention prediction is performed via ParRMDN, a
task-dependent ensemble of recurrent mixture density networks
2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
Scene
information
Front camera
Maneuver Tracking
HAMMER
Driving Inputs
Steer
Throttle
Vehicle State
Location
Orientation
Speed
Visual
Features
Extraction
C3D
Visual Attention
Task-Oriented
Attention
Allocator
ParRMDN
Fig. 1. Overview of HammerDrive. The proposed architecture uses scene and
telemetry data (red). HAMMER (blue) uses telemetry data to track the current
driving maneuver in real-time. A front camera provides scene information to a
convolutional 3D (C3D) network, a feature extraction module. These features
are processed by ParRMDN (green), an ensemble of neural networks, whose
outputs are scaled according to the HAMMER task feedback signal.
(RMDNs) [
26
]. In HammerDrive, HAMMER recognizes in
real-time the maneuver the driver is currently performing and
sends this information to ParRMDN, which predicts the focus of
attention of the driver. Although we investigate the application
of the proposed framework in simulated driving, we advocate
that the highly modular nature of HAMMER allows our model
to be generalized also to non driving-related applications. The
proposed architecture is trained and evaluated on data acquired
from our custom driving simulator.
The contributions of the paper are as follows:
1)
We present a novel architecture for real-time driver focus
prediction, where multiple visual attention modules are
dynamically activated via a task-driven attention scheduler;
2)
We evaluate HAMMER task-monitoring performance
over three different driving maneuvers, namely lane
maintenance, lane changes to the left and right, and
investigate performance-complexity trade-offs of data-
driven and model-based implementations;
3)
We analyze and compare the proposed task-aware visual
attention architecture against RMDN in terms of infor-
mation gain, Kullback-Leibler divergence, similarity and
cross correlation.
The rest of the paper is organized as follows: Section II
provides an overview of related works. Section III formalizes
the problem of visual attention prediction and introduces the
proposed architecture. Section IV focuses on an in-depth
explanation of HAMMER and its implementation for simulated
driving, while Section V describes both implementation and
training procedure of the RMDN modules. The data acquisition
methodology is described in Section VI, highlighting both the
experimental setup and data collection process. Section VII
analyses the results and performance achieved by the proposed
architecture. Section VIII discusses limitations of HammerDrive
and lists open challenges. Finally, Section IX summarizes
the contributions of the paper and outlines future research
directions.
II. RE LATE D WOR KS
HammerDrive is related to computational visual attention
modeling in driving (Section
II-A
) and biologically inspired
driver behavior modeling for Advanced Driver-Assistance
Systems (Section II-B).
A. Visual Attention Modeling in Driving
Driving represents a key application of visual attention
modeling in top-down driven daily tasks [
11
]. We identify
two main approaches to infer driver visual attention: one
where it is estimated via an interior camera facing the driver
exclusively [
27
], [
28
], [
29
], and one where it is combined with
an additional camera facing the scene [
5
], [
6
], [
30
], [
31
], [
32
],
[33].
The works in [
27
], [
28
], [
29
] exploit interior cameras facing
the driver to estimate with high precision whether the driver is
focusing their attention at the rear-view mirror, windshield or
dashboard. However, such approaches cannot infer the focus
of attention within the scene in front of the driver [
27
], [
29
],
but only whether the driver is looking through the windshield.
To estimate if the driver is looking at a crossing pedestrian or
traffic lights, a secondary camera, such as the one assumed for
HammerDrive or maneuver assistance [22], [34], is needed.
In line with this, Palazzi et al. in [
5
] introduced a multiple
bottom-up branch model that employs Convolutional Neural
Networks (CNNs) and leverages on visual information from
the scene, motion from optical flow and semantic segmen-
tation to refine fixation maps. Similarly, [
30
] introduced a
fully CNN-based network for human attention prediction in
driving video observation. These models achieve state-of-the-
art performance in video saliency detection in driving, however
they are inherently task-unaware. Sensory based, i.e., bottom-
up, approaches are limited in visual attention modeling, as
past literature showed that human attention patterns greatly
depend on the task and sub-tasks, especially in driving [
35
].
Recent task-aware top-down approaches have proven to be
very promising for visual attention modeling in driving, but
they have been tested on datasets where observers were either
looking at static images [
6
], [
32
] or not actively driving [
31
].
However, human vision and actions are strongly correlated, as
gazes are used to collect the information required to perform
an action [
36
]. Toward this end, HammerDrive integrates active
real-time maneuver tracking and exploits this information to
guide visual attention prediction towards meaningful goals for
the driver in a top-down manner.
B. Biologically Inspired Driver Behavior Modeling for ADAS
The proposed framework for visual attention prediction is
also closely related to driver modeling for Advanced Driver-
Assistance Systems (ADAS). Driver modeling can be applied to
a large number of problems; here, we focus on the challenging
task of building human-like models of expert drivers to
intervene in a shared-control manner.
Building on the concept of artificial co-driver [
37
], authors
in [
38
], [
39
] have designed biologically inspired multi-layered
systems that replicate expert human behavior to provide shared-
control assistance in driving. Similarly to HammerDrive, these
works build on modular hierarchical systems that perform goal-
oriented action prediction. The system simultaneously computes
action request signals related to all plausible goals, and then
sorts them according to a fitness criterion to perform take-
over maneuvers. Authors in [
40
] have developed an assistance
system that combines driver modeling to mirror human-like
AMADORI et al.: HAMMERDRIVE: A TASK-AWARE DRIVING VISUAL ATTENTION MODEL 3
Clip c(t)
Vehicle State
Visual Scene
clip
System State
x-position
y-position
x-forward
y-forward
speed
Input State
steer
throttle
Error Estimation
Weighted Normalized Sum
Left Lane Change
RMDN
Lane Keep
RMDN
Right Lane Change
RMDN
Confidence Instance
Vector
Confidence Level
Vector
Inverse
Model
Inverse
Model
Left Lane Change
Waypoint Generator
Lane Keep
Waypoint Generator
Right Lane Change
Waypoint Generator
Forward
Model
Forward
Model
Forward
Model
Inverse
Model
System State
s(t)
τ
x
x-position
τ
y
y-position
θ
x
x-forward
θ
y
y-forward
v
speed
s(t+1) ψ (t+1)
Final Driver Focus Prediction
C3D Network
ψ (t+1)
C
64
M
P
C
128
M
P
C
256
C
256
M
P
M
P
C
512
C
512
M
P
C
512
C
512
ξ(t) P(l,t)
P(t)
e(t+1)s(t+1)
^
s(t+1)
^
Fig. 2. Visual attention inference procedure in HammerDrive. The block to the left (red) lists the sensor readings required during validation. The bottom
blocks (blue) show task tracking using HAMMER network via easy-to-access telemetry data. The top network (green) depicts both C3D and ParRMDN and
represents the visual attention predicting component of HammerDrive. The outputs from both networks are combined as a weighted-normalized sum (purple).
maneuvers with driver monitoring to detect drowsiness and
inattention to trigger take-overs from the system.
The proposed HammerDrive architecture is also closely
related to the Adaptive Control of ThoughtRational (ACT-R)
driver model from [
41
], [
42
], which has been successfully
applied to describe driver behavior in a vast number of
scenarios. The model builds on the ACT-R architecture [
43
],
which exploits on modularity, seriality and parallelism to infer
driver behavior and focus of attention during driving. The
ACT-R model comprises of three primary components: control,
monitoring and decision making. The control component
directly relates to HammerDrive, as it regulates the relationship
between perception and vehicle manipulation for lateral and
longitudinal control, i.e., steering and accelerating, respectively.
Driver modeling can also be applied to autonomous vehi-
cles [
44
], as human-like behavior is often considered safer
and more acceptable for users [
45
]. In line with the above,
the proposed framework defines a human-like computational
model of visual attention, by building on biological evidence
of action planning methodologies in humans [25].
III. PROB LE M AN D MOD EL
We formalize driver visual attention modeling as a special
case of eye gaze fixation prediction. Eye fixation prediction is
defined as a general problem of function estimation.
A. Problem Formulation
Given a group of
Nsub
subjects, we define the corresponding
visual attention dataset as
D=n(xs
i|cs
i)N
i=1oNsub
s=1 ,(1)
where the tuple
(xs
i|cs
i)
identifies two
tf
-long sequences:
a) a sequence of gaze locations
xs
i= [xs
i(t)]tf
t=0
and b) a
sequence of frames
cs
i= [cs
i(t)]tf
t=0
, which we refer to as a
clip. Sub-indices
i
and super-indices
s
differentiate between
data samples and subjects, respectively. In other words,
xs
i(t)
is the instantaneous gaze location of the
s
-th subject while
they were looking at frame cs
i(t).
The task of visual attention modeling is the derivation of a
function
f(·)
, that can estimate the most likely gaze location
of a driver
ˆxs
i(t)
given a specific frame
cs
i(t)
. This corresponds
to the following optimization problem
P: min
f∈F
Nsub
X
s=1
N
X
i=1
Γ (f(cs
i),xs
i),(2)
where the operator
Γ(·)
is used to identify the chosen loss
function, in our case the negative log-likelihood, see Section V.
In the following subsection, we introduce the proposed model,
namely HammerDrive, to solve the optimization problem P.
B. HammerDrive Model
The proposed HammerDrive model, visually presented in
Fig. 2, is characterized by two main networks: ParRMDN, an
ensemble of Recurrent Mixture Density Networks (RMDN) for
visual attention prediction (top, green) and a HAMMER-based
action recognition network (bottom, blue). Our architecture
builds upon the concept that human visual attention can be
modeled as an ensemble of bottom-up networks competing
for resources, given top-down task-aware guidance. Under
this assumption, the competition for resources is modeled and
controlled via a task-aware top-down driven attention network.
1) HAMMER
acts as a top-down driven attention network
that recognizes in real-time the maneuver of the driver and
instructs the ParRMDN network on which modules to activate.
The network can be formalized as an ensemble of multiple
parallel modules [
25
], each designed to provide the confi-
dence/likelihood of a specific driving maneuver. The HAMMER
network offers great flexibility in the implementation of
its modules, allowing the coexistence of modules based on
kinematic models, neural networks or Kalman filters, see [
46
].
After the confidence values of all modules have been computed,
they are used to perform a weighted sum of ParRMDN
predictions. We describe HAMMER in Section IV.
2) ParRMDN
operates as an ensemble of RMDN modules,
each trained to predict visual attention in the occurrence of a
specific driving maneuver. The modules are defined according
to [
26
] and build upon the concept of cascading a recurrent
neural network with mixture density networks [
47
]. We perform
feature extraction over a
tf
-long clip of images via a pre-
trained Convolutional 3D (C3D) network [
48
]. We introduce
the formulation and training of RMDN modules in Section V.
4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
IV. HAMMER
Task-awareness in HammerDrive is achieved by means
of a HAMMER-based action recognition network [
25
]. The
network recognizes a set of actions (maneuvers)
L
with
cardinality
kLk =L
and consists of
L
pairs of inverse and
forward models operating in parallel. This section provides a
detailed introduction to HAMMER’s paradigm, together with
its pseudocode implementation in Algorithm 1.
A. Overview
Considering a maneuver
l∈ L
, we define the corresponding
inverse model
fl
I
as the function that takes as input the system
state
s(t)
at time step
t
, and computes the control commands
al(t)
to be applied to the system to achieve that goal. The
forward model
fl
F
takes as input the current state of the system
s(t)
and control commands
al(t)
to compute the predicted
state of the system at the next time step
ˆ
sl(t+ 1)
. Once
the predicted state of the system for the
l
-th module has
been computed by sequentially evaluating inverse and forward
models, it is then compared to the actual system state at the next
time step
s(t+ 1)
. The comparison results in an error signal
that is used to increase or decrease the confidence value of the
maneuver that corresponds to the l-th module. The maneuver
corresponding to the module with the highest confidence value
is considered as an estimate of the driver’s intention.
In our study, we focus on
L= 3
maneuvers: 1) lane
maintenance, 2) lane change to the left and 3) lane change
to the right. Since all these maneuvers relate to control inputs
over the steering wheel, the definition of corresponding inverse-
forward model pairs, i.e., the cascade of
fl
Ifl
F,l∈ L
,
is equivalent. Although analytically equivalent, each inverse-
forward model pair differs in the tangential angle assumed for
the way-point. When defining the inverse and forward models
for driving maneuvers, we assume the vehicle to be moving
on a R2workspace W.
We define the state vector s(t)in time tas
s(t) := [τx(t), τy(t), θx(t), θy(t), v(t)] ,(3)
where
[τx(t), τy(t), v(t)]
, represent the
x
-axis,
y
-axis location
of the car and speed, respectively, and
[θx(t), θy(t)]
identify
the
x
and
y
-axis components of the unitary forward vector of
the car, i.e., the car heading.
Similarly, we define the input state vector u(t)as follows
u(t) := [w(t), g(t)] ,(4)
where
w(t)[wmax, wmax ]
represents the steering wheel
angle, with
wmax
being the maximum turning angle, and
g(t)
[0, gmax]
is the throttle pedal position of the car, with
gmax
being the maximum allowed. We normalize both steering and
pedal range using a simple min-max normalization, hence
w(t)[1,1] and g(t)[0,1].
B. Inverse Model Formulation
We formulate the inverse models of HAMMER as functions
that compute the steering input required to move the car from its
current state
s(t)
towards the
L
possible behavioral locations.
Algorithm 1 HAMMER Algorithm
Input: State vectors s(t)and s(t+ 1)
Output: Recognized maneuver l
Receive new observation from system s(t+ 1)
for l∈ L do
Compute way-point direction φl
Compute steering input ˆw(φl)via Eq. (5)
Compute predicted state input ˆ
s(t+ 1) via Eq. (6)
Compute error value e(t+ 1) via Eq. (8)
end for
Identify most probable module as arg min e(t+ 1)
Compute confidence instance ψ(t+ 1) via Eq. (9)
Compute confidence level vector ψ(t+ 1) via Eq. (10)
return Identify recognized maneuver as arg max ψ(t+ 1)
We employ an orientation correction approach [
49
], where
we discretize a unidirectional road into three separate virtual
lanes [
4
]. Given the current state of the car
s(t)
we define three
way-points or maneuver goals, one per behavioral location and
located at fixed road-distance from the location of the vehicle.
After the way-points have been identified and their location has
been computed, the algorithm evaluates the angular distance
between the direction of the car in its current state, i.e.,
[θx, θy]
,
and the tangential angle of the
l
-th way-point
φl
. The angular
differences identify the set of steering angles required for the
car to proceed towards the corresponding way-points and is
used to compute an estimate of the steering input
ˆw(t)
via linear
mapping. We define the inverse model function for steering as
fl
I: ˆw(φl) = (θφl)
wmax
=(arctan (θyx)φl)
wmax
,(5)
where
θ
represents the direction of the car or yaw and
φl
identifies the tangential angle for the
l
-th way-point. Time-
dependence of parameters has been removed to ease notation.
C. Forward Model Formulation
The forward model computes the expected behavior of the
vehicle, given its current state and inputs. The formulation
of a forward model is inherently a trade-off between accu-
racy and complexity. In our case, we favor simplicity over
accuracy, as the proposed architecture is required to rapidly
and simultaneously compute the expected behavior of the
vehicle for multiple maneuvers. We assume a kinematic bicycle
model, which was shown to achieve good accuracy when
modeling vehicle behavior in real-driving scenarios, despite its
simplicity [
50
]. Under these assumptions, if the time interval
between observation
δt
is small, the car moves in the same
direction of its rear wheels, i.e., the car direction
(θx, θy)
in
our case. Following the same notation as for the inverse model,
we formulate the predicted state of the car for the
l
-th goal as
fl
F:ˆ
sl(t+ 1) =
ˆ
θx= cos(θ+ ˆwl·wmax )
ˆ
θy= sin(θ+ ˆwl·wmax )
ˆv=v0+ ˙v·δt=v0+a·δt
ˆτx=x0+ ˙x·δt=x0+ ˆv·cos(θ)·δt
ˆτy=y0+ ˙y·δt=y0+ ˆv·sin(θ)·δt
.(6)
While we presented a kinematics-based approach for
HAMMER, its high modularity allows the inclusion of more
complex models, such as inverse-forward pairs of Neural Net-
work (NN) models, which we will evaluate in Section VII-C.
AMADORI et al.: HAMMERDRIVE: A TASK-AWARE DRIVING VISUAL ATTENTION MODEL 5
D. Confidence Extractor
Once inverse-forward model pairs have been computed for all
the considered maneuvers, their outputs are used at a prediction
verification stage to generate a set of error signals. We define
the RL×1error vector e(t+ 1) as
e(t+ 1) = [el(t+ 1),l∈ L]
= [(ˆ
sl(t+ 1),s(t+ 1)),l∈ L],(7)
where
el(t+1)
,
ˆ
sl(t+1)
and
s(t+1)
identify the error value, the
predicted state and the actual state corresponding to the
l
-th ma-
neuver, respectively. Here, the operator
(ˆ
sl(t+ 1),s(t+ 1))
is used to represent the error function that compares the state
vectors received in argument. Without loss of generality, we
compute
(·)
as the Euclidean distance between the predicted
forward direction vector of the vehicle for the
l
-th maneuver
[ˆ
θx,l,ˆ
θy,l]and its true direction. Therefore, we have
(ˆ
sl(t+ 1),s(t+ 1)) = q(ˆ
θx,l θx)2+ (ˆ
θy,l θy)2.(8)
The error vector
e(t+1)
is then used to compute a confidence
instance vector
ψ(t+ 1)
. The confidence instance is a
L×1
vector whose elements are all zeros, except for the element
that corresponds to the goal with the lowest error value, which
is given a unitary value:
ψl(t+ 1) = (0if l6= arg min e(t+ 1)
1if l= arg min e(t+ 1) .(9)
Confidence instances are then collected in a confidence level
vector
ψ(t+ 1)
according to a confidence window of length
cw
. We define the confidence level vector as the weighted sum
of the cwprevious confidence instances:
ψ(t+ 1) =
cw
X
i=0
aiψ(t+ 1 i),(10)
where
ai
identifies the time-based weights for the confidence
instances. In our implementation, unless differently specified,
we assume a confidence window
cw= 1s
as it shows a positive
trade-off between real-time predictions and performance (see
Section
VII-C
), and we consider a simple linear weighting
scheme for confidence instances where recent instances have
higher influence for task recognition than past ones, i.e.,
ai=
(cwi)/cw.
Finally, the task corresponding to the highest element in
the confidence level vector is identified as the maneuver that
the driver is currently performing. This signal is propagated
through the network and is used to perform a weighted sum
of the visual attention prediction outputs of ParRMDN, as
shown in Fig. 2. More details on visual attention networks are
provided in the next section.
V. RMDN MODULE
This section introduces the modules of ParRMDN. The
network is an ensemble of RMDN networks following a
single C3D feature extraction network [
26
], [
48
], see Fig. 3.
The feature vectors are then input to the
L
modules, each
characterized by a cascade of a long short-term memory
network (LSTM) [51] and a mixed density network (MDN).
C3D
LSTM
MLP
MDN
GMM
Loss
C3D
LSTM
MLP
MDN
GMM
C3D
ct-T ct
ξt-T
ht-T ht
yt-T yt
pt-T pt
Shared
Weights
RMDN RMDN
ct-T ct
P(t) ψ(t)xt
ξt
Fig. 3. ParRMDN module block diagram. During training (left), a
T
-long
sequence of
16
frames clips
c={ctT, ..., ct}
is input to a C3D network
(pre-trained on Sports-1M), which outputs a sequence of feature vectors
x={xtT, ..., xt}
. The sequence is fed to a LSTM network, whose hidden
states
ht
are projected via multilayer perceptron (MLP) to a vector of Gaussian
parameters
pt
. The loss is the negative log-likelihood of the ground-truth gaze
location against the mixture of Gaussians
pt
via Eq. (19). During inference
(right), each clip of a
T
long sequence is first processed using a shared C3D
network. The sequential output
x
is then simultaneously fed to the RMDN
modules, each providing a maneuver-dependent bivariate Gaussian prediction.
The predictions are then combined according to Eq. (21).
A. Model Formulation
We assume a dataset with structure
D=vi,xi, liN
i=1
,
where
N
identifies the number of available triples data-points.
Here,
vi
represents the video data,
xi
is used to represent
the ground-truth gaze locations and
li
the associated sub-task.
Video data
vi
comprises
T= 14
overlapping clips
ci
t
of
tf= 16
frames, i.e.,
vi=ci
tT1
t=0
, which corresponds to
scene information spanning
T+tf= 30
frames, i.e.,
1s
. Gaze
data points
xi
are tuples of
(x, y)i
positions, normalized to
[0,1].
Given the
i
-th datapoint of
D
, RMDN first extracts the
C3D features of the input clip as
ξt=C3D(ci
t)
. This
operation is performed independently from the task, i.e.,
ParRMDN only requires a single C3D network. C3D is defined
as [
48
]: C64-MP-C128-MP-C256-C256-MP-C512-C512-MP-
C512-C512-MP, where C represents a three-dimensional con-
volutional layer, MP is the max-pooling layer and the number
specifies the number of kernels of the layer. The first MP layer
has kernel (1, 2, 2), whereas all others have a (2, 2, 2) kernel.
Once the vector of features
ξt
has been computed, the LSTM
network operates as follows:
it=σ(Wiht1+Iiξt+bi)(11)
ft=σ(Wfht1+Ifξt+bf)(12)
ot=σ(Woht1+Ioξt+bo)(13)
ct=ftct1+ittanh(Wcht1+Icξt+bc)(14)
ht=ottanh(ct),(15)
where
it
,
ft
,
ot
,
ct
and
ht
identify input gate, forget gate, output
gate, memory cell and hidden representation, respectively.
Operators
σ
and
tanh
represent the element-wise sigmoid
function and hyperbolic tangent function, respectively. The
6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
parameters to be learned are
W
,
I
and
b
, where
is used
to represent {i, f, o}.
The hidden representation
ht
of the LSTM network is input
to a linear layer that projects the parameters to a mixture of
C
bivariate Gaussians to compute the two-dimensional visual
attention map. Given a re-parameterization of the model as
a Gaussian Mixture Model (GMM) with
C
components, we
have
yt=Wyht+by={πc
t,˜µc
t,˜σc
t,˜ρc
t)}C
c=1 (16)
where
Wy
and
by
are the weights and bias of the projecting
linear layer, respectively. Since the outputs of the linear layer
are unbounded real numbers, we normalize their values to de-
fine a valid probability distribution
pt={(µc
t, πc
t, σc
t, ρc
t)}C
c=1
as follows [47]:
pt=
πc
t=exp(˜πc
t)
PM
c0=1 exp(˜πc0
t)
µc
t= ˜µc
t
σc
t= exp(˜σc
t)
ρc
t= tanh(˜ρc
t)
C
c=1,
(17)
where
µc
t
and
σc
t
are means and standard deviations of the
bivariate Gaussians, respectively,
πc
t
are the mixing coefficients
with
PC
c=1(πc
t)=1
and
ρc
t
identify the correlations between
variables.
B. Model Training
The only trainable components of HammerDrive are the
modules of ParRMDN. We train the modules individually on
behavioral-aware sub-datasets
Dl
, which we define as subsets
of D, i.e., Dl⊂ D ,l∈ L. Analytically we have
Dl=vi,xi, liNl
i=1 , s.t. li=li0, ..., Nl,(18)
where
Nl
identifies the number of data-points available in the
sub-dataset corresponding to the
l
-th maneuver. For each
Dl
,
80%
of the sub-dataset is used for training and the remaining
20% is used for testing.
The training of ParRMDN modules is performed by opti-
mizing the negative log-likelihood of the ground-truth gaze
locations xifor the i-th frame sequence vi:
Γ(vi,xi) =
T1
X
t=0
A
X
j=1 log C
X
c=1
πc
tN(ai
t,j ;µc
t, σc
t, ρc
t)!.
(19)
For the interested reader, more details on RMDNs and their
training can be found in [47].
C. Task Awareness Integration
We have seen that ParRMDN operates as an ensemble
of RMDN modules, each trained to predict the maneuver-
dependent visual attention of the driver. At inference stage, all
modules produce different predictions from the same video data
and they are averaged according to the confidence level vector
given by HAMMER (see Eq.(10)). Given the instantaneous
prediction of a single RMDN module
P(l, t) =
C
X
c=1
πc
tN(µc
t, σc
t, ρc
t),(20)
Fig. 4. Driving simulator setup. Top: simulated environment with an overlay
of the three virtual lanes considered, their width and the free-flow speed. Note
that none of these overlays were visible to participants during the experiment.
Bottom: the participant wears a VR headset with integrated eye tracker. The
screen shows the scene displayed to the participant and real-time sensor
readings for monitoring purposes.
the output of HammerDrive is computed as:
P(t) = hψl(t+ 1) ·P(l, t)iL
l=1 .(21)
VI. EXPERIMENTS
We evaluate HammerDrive on a dataset that we collected on
our custom designed driving virtual reality (VR) simulator (see
Fig. 4). The experiment was designed to collect physiological
signals that were unobtrusive, namely gaze locations from a
VR headset, and easy-to-access telemetry data from the vehicle,
namely steering angle, throttle, speed, location and scene
information. This study has been approved by the Ministry of
Defence Research Ethics Committee (MoDREC).
A. Participants
We recruited twenty participants (mean age 24.2, standard
deviation 4.5), all experienced drivers with normal or corrected
vision. Before the trial, participants were introduced to the
sensors, experimental setup and a brief explanation of the task
to be performed. The eye-tracking sensors were calibrated
at the beginning of each trial. To avoid learning effects,
each participant was allowed a demo trial to familiarize
themselves with the driving simulator before performing the
actual experiment.
B. Setup
We set up a realistic driver-in-the-loop simulation for the
experiment (see Fig. 4). The setup comprised of a physical
simulator, a VR headset with integrated eye gaze tracking,
which required an infra-red camera mounted above the steering
wheel and a screen to monitor the participants. The simulated
driving environment was custom developed and designed
using the Unreal Engine (https://www.unrealengine.com). The
AMADORI et al.: HAMMERDRIVE: A TASK-AWARE DRIVING VISUAL ATTENTION MODEL 7
engine is known to provide state-of-the-art near photo-realistic
rendering quality and cutting edge realistic physics, and has
been successfully used in the past literature for a range
of applications, including realistic driving simulators for
autonomous vehicle research [52], [53].
C. Experimental Procedure
The experiment required each participant to drive for five
minutes along a straight
10
m-wide highway on our simulator.
Multiple rectangular shaped obstacles are placed on the track,
and the participants are asked to avoid these obstacles while
driving at a free-flow speed of
120
km/h. For the purpose of
simulation, the road is discretized into three lanes with
3.3
m
width and the obstacles are randomly placed at one of three
lanes at a fixed distance between each other. The obstacles are
3.5
m wide, so that they entirely block a lane in width, and they
are
120
m distant from each other. Therefore, given a speed
of
120
km/h, participants pass an obstacle approximately every
3.5
seconds. The total length of the highway is
15
km, however
participants were asked to drive five minutes per session, which,
at a speed of 120km/h corresponds to 10km.
The experiment and the simulated scenario are designed
to reduce inter-participant differences in the rate of driving
maneuvers, the environment experienced and the cognitive
states. We achieve this by maintaining all participants’ vehicles
to drive at a constant velocity, i.e.,
120
km/h, even when
avoiding the obstacles. While there are no theoretical constraints
to implementing braking and turns in HammerDrive, we restrict
driving maneuvers to lane changes. This ensures that all drivers
perform the same number of driving maneuvers, and that
each maneuver is performed in the same amount of time,
i.e., every
3.5 seconds. Besides ensuring that all drivers
experienced similar levels of cognitive states and engagement,
the chosen scenario, i.e., a highway driving with lane-changes,
also ensures that the collected dataset captures a richer set of
human behaviors for each of the considered maneuvers.
To limit situations where drivers can avoid multiple subse-
quent obstacles without performing a lane change, we employ a
custom discrete distribution for obstacle lane placement. Given
the number of obstacles between the current obstacle and the
previous one on the
i
-th lane
di
, we define the next obstacle
placement probability distribution on lane ias
p(i) = edi
Piedi,
(22)
where
δ
identifies the distance between two adjacent obstacles.
This ensures that if the
i
-th lane has not been blocked for the
past
4
obstacles, the probability of blocking that lane is
e4
times higher than that of the most recently blocked lane. Note
that obstacle placement is designed such that obstacles only
cover one virtual lane, i.e., two of the three virtual lanes are
empty so that drivers are always able to avoid the obstacles.
D. Dataset Collection
For each participant, we collected (see Section
IV-A
):
1) instantaneous two-dimensional gaze locations for both left
and right eye (
{ts, xr, yr, xl, yl}
), 2) state information of the
TABLE I
SUM MARY O F DATASE T FEATU RE S.
Dataset Frames Subj. Annotations Active
HammerDrive 180,000 20 GMap, Tlmy, DrInp Yes
Deng et al. [31] 74,825 28 GMap No
Palazzi et al. [5] 555,000 8 GMap, Tlmy Yes
Pugeault & Bowden [54] 158,668 1 GMap, Tlmy, DrInp Yes
vehicle (
{τx, τy, θx, θy, v}
), 3) driver inputs (
{w, g}
) (all at
60 Hz) and scene observed (
200 ×112
sized RGB images at
30Hz). The integrated eye-tracker in the VR headset provides
gaze data as a set of two-dimensional coordinates, which
correspond to the human gaze location observed on the left
and right eye screens, as seen through the headset lenses. The
coordinate system of the screens employed by the VR headset
is normalized to the range
[1,1]
along both x-axis and y-axis,
so that its origin is
(0,0)
, bottom-left corner is
(1,1)
and
top-right corner is
(1,1)
. Each participant performed a drive
of 5 minutes, while their gaze, the scene they observed and
their behavioral telemetry were recorded.
For our dataset, we recorded a total of 9000 frames and
18000 samples per data stream per participant. We recap the
features of our dataset in Table I and compare it with related
datasets. The table indicates the total number of frames, the
number of subjects, the annotations included and whether the
subjects were actively driving during the experiments. Here,
GMap, Tlmy and DrInp indicate gaze maps, telemetry and
driver inputs, respectively.
VII. RES ULTS
In this section, we present and discuss the performance of
the proposed task-aware visual attention model for driving,
namely HammerDrive. Results are computed using a 5-fold
validation and results are averaged over 20 realizations.
We compare the proposed model against the state-of-the-art
in deep learning based visual attention modeling, RMDN [
26
],
[
5
]. To the best of the authors’ knowledge, there is only a lim-
ited number of deep learning based video saliency models [
11
],
[
55
] to date, and RMDN is still referenced as a state-of-the-art
performer when evaluated on the challenging HOLLYWOOD-
2 dataset, [
26
], [
56
], [
55
]. We first show the performance of
task-aware ParRMDN modules; these results correspond to the
performance we would achieve under the assumption that an
oracle is providing the task currently being performed by the
driver (Section
VII-B
). We then show the performance achieved
by HAMMER in tracking the driving maneuvers as a function
of the number of available data-points and its robustness to
data-driven and model-based implementations (Section
VII-C
).
Finally, we show the performance of the proposed model
when both HAMMER and ParRMDN are jointly operating
(Section
VII-D
). We evaluate the performance of the proposed
scheme in terms of multiple visual attention metrics, i.e.,
Kullback-Leibler divergence (
KL
), cross-correlation (
CC
),
similarity (
Sim
) and information gain (
IG
) [
57
]. We describe
each of the chosen metrics in the following subsection.
A. Evaluation Metrics
The metrics are chosen as they all provide different insights
on the prediction provided by the model. In order to com-
8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
pute these metrics, we assume both predicted locations and
ground-truth gaze patterns to be two separate two-dimensional
distributions of probabilities, i.e.,
P
for model prediction and
Qgt for ground-truth.
KL
is a general measure that evaluates the difference
between two probability distributions, analytically
KL(P, Qg t) = X
i
Qgt
ilog +Qgt
i
+Pi!,(23)
where
Pi
identifies the sum over all the pixels and
is a
regularizing factor. Therefore, lower values indicate that the
two densities are more similar.
The CC metric computes how correlated P and Qgt are:
CC (P, Qgt ) = cov (P, Qgt)
cov(P)×cov(Qg t),(24)
where the
cov(·)
operator identifies the covariance. Clearly,
higher values of
CC
indicate that the two densities follow a
similar behavior.
The
Sim
metric numerically evaluates the similarity between
two distributions, analytically
Sim(P, Qgt) = X
i
min(Pi, Qgt
i).(25)
As it follows,
Sim
values are constrained between [0,1],
where 0 corresponds to highly dissimilar distributions and
1 to perfectly identical distributions.
Finally,
IG
is computed as the gain introduced by the
predicted distribution when compared to a systematic bias,
in our case a centered prior Gaussian, and the ground-truth
distributions. Analytically, it is computed as
IG(P, Qb) = 1
NX
i
Mgt
i[log2(+Pi)log2(+Bi)] ,
(26)
where
Mgt
is a binary map of gaze locations, and
B
is the
distribution of the chosen bias.
B. ParRMDN Modules
Table II includes the visual attention inference performance
for each of the ParRMDN modules, under the assumption that
an oracle is providing the correct task currently performed. As
mentioned in Section
V-B
, ParRMDN modules are trained on
the training portion of the corresponding maneuver sub-datasets
(i.e.,
80%
of
Dl,l∈ L
). To highlight the benefits of the
proposed task-aware module training, we collect performance
of each module both on sequences corresponding to the
same maneuver the module was trained for and on sequences
corresponding to different maneuvers. For instance, for the left
lane change module,
Dleft
identifies the performance on left lane
change dataset, while sequences that belong to other maneuvers
are presented as
D6=left =Dkeep ∪ Dright
. Performance of each
module are averaged over
25
test sequences of
25
frames each.
As shown in Table II, task-aware module performances
are always higher when evaluated on test sequences from the
corresponding task-aware sub-dataset, e.g., left module on
Dleft
,
whereas performance decrease when the module is tested on
different task sub-datasets
D6=left
. Performance gaps for
KL
,
TABLE II
TASK -DE PE NDE NT MODULE PERFORMANCE
Mod. Left Mod. Keep Mod. Right No Task RMDN
Dleft D6=left Dkeep D6=keep Dright D6=right D D
KL 0.72 0.91 0.70 0.78 0.71 0.84 0.78 0.81
CC 0.72 0.63 0.72 0.65 0.73 0.64 0.69 0.69
Sim 0.56 0.48 0.55 0.49 0.56 0.48 0.52 0.48
IG 4.00 3.22 3.16 2.89 5.49 3.15 4.02 3.76
Metric where higher values indicate better performance.
Metric where lower values indicate better performance.
CC
,
Sim
show that task-aware modules are able to better
predict the focus of the driver when the task being performed
is its corresponding one.At the same time, large gaps on
IG
show that predictions on corresponding tasks are considerably
more meaningful than the ones during different tasks.
Table II also includes the performance achieved by task-
unaware HammerDrive, identified by No Task, and RMDN.
For task-unaware HammerDrive, we set the confidence level
vector to
ψ(t) = [0.33,0.33,0.33]
. This corresponds to a case
where no information on the maneuvers is available, hence
HammerDrive assumes all modules to have equal weights at the
weighted sum stage. As we can see, task-based modules achieve
higher performance than task-unaware HammerDrive, proving
that the performance gap is motivated by task-awareness. On the
other hand, task-unaware HammerDrive shows comparable and
for some metrics better performance than RMDN. These results
are not surprising, as task-unaware HammerDrive corresponds
to an ensemble of RMDN modules, each trained on a specific
portion of the dataset. The performance gap between task-aware
modules, task-unaware HammerDrive and RMDN confirms the
findings from the literature [
5
], [
35
], [
41
], which showed that
gaze behavior of drivers is indicative of their current maneuver.
C. Robustness to Data-Driven and Model-Driven approaches
Given the flexibility of the HAMMER formulation, we can
follow either a model-driven or a data-driven approach during
forward-inverse model pairs design. In this section, we compare
two implementations of HAMMER in terms of model accuracy
and task recognition accuracy as a function of confidence
window length
cw
. Kin identifies the kinematics-based model-
driven implementation (see Section IV), whereas NN represents
the NN-based data-driven approach.
In the NN approach, we define forward and inverse models
as single multilayer perceptron feed forward networks with
a single hidden layer of 8 units. Forward and inverse model
networks have been trained on the telemetry data of a single
participant and are then tested on unseen participants’ data.
Model accuracy is shown in Figs. 5a and 5b, for inverse
and forward model, respectively. In Fig. 5a, we compare the
true steering input
w(t+1)
with its estimated value
ˆw(t+1)
from both Kin and NN inverse models. In a similar manner,
Fig. 5b shows a comparison between the true direction vector
[θx, θy](t+1)
and its estimation from the Kin and NN forward
models, i.e.,
[ˆ
θx,ˆ
θy](t+1)
. Both figures show that a data-
driven approach provides a more accurate representation of the
vehicle kinematics, when compared to an analytic model-driven
approach. Both models offer accurate representations of the
1-step vehicle behavior, which is a fundamental requirement
AMADORI et al.: HAMMERDRIVE: A TASK-AWARE DRIVING VISUAL ATTENTION MODEL 9
(a) Inference of steering input ˆw.(b) Inference of vehicle heading ˆ
θ.
Fig. 5. Inverse and forward module inference performance. NN identifies
prediction from shallow data-driven model, while Kin stands for the prediction
achieved via Kinematic model from Eq.(5) and Eq.(6) for inverse and forward
module, respectively. Vehicle heading is computed as ˆ
θ= arctan(θyx).
0.5 1.0 1.5 2.0 2.5 3.0
Confidence Window Length [s]
0.5
0.6
0.7
0.8
0.9
1.0
Recognition accuracy
Kin Lane Change Tasks
Kin Lane Keep Task
NN Lane Change Tasks
NN Lane Keep Task
Fig. 6. HAMMER task recognition accuracy as a function of the length of the
confidence window. NN identifies data-driven implementation of HAMMER,
while Kin stands for a kinematic model implementation.
for HAMMER, as it ensures that both forward and inverse
models predictions are reliable for maneuver tracking.
Although the performance shown in Fig. 5 are promising,
we are more interested in how the two implementations affect
HAMMER in its ability to correctly track maneuvers on
a longer prediction horizon. In Fig. 6 we show the task
recognition accuracy of NN-based and Kin-based HAMMER.
The NN-based implementation shows higher task recognition
accuracy during active tasks, namely lane changes on both
left and right lane, while lane maintenance suffers by lower
accuracy, due to jerky movements that often happen during
lateral control. These movements can be mistakenly interpreted
by a very accurate model, such as NN. On the other hand, these
events are filtered by Kin-based HAMMER, whose accuracy is
95% in both scenarios with
cw= 1s
, i.e., the value considered
for HammerDrive. Since each maneuver instance normally lasts
3.5s
, these results indicate that HAMMER can reliably track
and recognize a maneuver
2.5s
in advance, or before, if we
assume shorter cw.
D. HammerDrive Performance
We collect HammerDrive performance in Fig. 7 for all
presented metrics. Results show that task-awareness is greatly
beneficial to achieve reliable visual attention predictions, with
clear gains over standard RMDN. Figs. 7a and 7c highlight that
HammerDrive prediction follows the actual focus of attention
of the driver more closely than RMDN, as the proposed
method outperforms the state-of-the-art on
KL
and
Sim
,
with improvements on performance of
13%
in both cases.
Similarly, Fig. 7d indicates that HammerDrive is able to provide
more meaningful predictions as achieved
IG
are
12%
higher
than the ones achieved by RMDN. Fig. 7b shows that both
methods are able to capture the dynamics of focus of attention,
as they both reach comparable
CC
values. It is important to
stress that HammerDrive is very consistent in its prediction
and performance in comparison with RMDN. As shown in
Fig. 7, all performance metrics for HammerDrive have lower
variation with smaller standard deviations.
We provide a qualitative assessment of the predicted visual
attention distribution in Fig. 8, which includes predictions for
all driving maneuvers considered in addition to a failure case.
It is interesting to notice that, on the failure case portrayed in
Fig. 8d, the driver was looking at the road in the immediate
surrounding of the vehicle, without paying much attention to
the obstacles ahead. If we compare this with HammerDrive’s
prediction, we can see that the model expected the driver to
be looking at the obstacles instead. Therefore, Fig. 8d shows
a case where HammerDrive could be applied in ADAS as the
driver’s focus is not on relevant areas of the scene.
Since the final goal of the proposed HammerDrive archi-
tecture resides in its ability to be implemented and operate in
ADAS, we evaluate its computational time during inference.
Given a clip of
16
frames, C3D feature extraction and Par-
RMDN inference only require
6
ms and
5ms, respectively. The
total inference time is
11
ms, which corresponds to an inference
rate of
90
Hz. Since the scene information is captured at
30
Hz, we can conclude that the proposed HammerDrive is able
to operate in real-time.
VIII. DISCUSSION, LIMITATIONS AND OPEN CHALLENGES
We introduced HammerDrive, a model for driver visual
attention prediction that dynamically integrates and exploits
task-awareness. Our model builds on the assumption that a
driver gaze pattern is driven by the goal of completing the
maneuver he is currently performing. This assumption has
been validated in the past literature by numerous studies and
models, such as the ACT-R driver cognitive model [
41
], [
43
].
In ACT-R a two-state process is assumed for lateral control: an
initial state for salience perception, where the driver shifts the
visual attention to compute the steering needed to achieve a
goal, and a motor state, wherein the estimated steering for the
chosen goal is acted on. In a similar manner, HammerDrive
uses HAMMER to compute all possible maneuvers and to
select the most probable. Once the most likely maneuver has
10 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
HammerDrive RMDN
Method
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
KL
s.d. = 1.16e-02
s.d. = 2.71e-02
(a) Box plot for KL metric.
HammerDrive RMDN
Method
0.600
0.625
0.650
0.675
0.700
0.725
0.750
0.775
0.800
cc
s.d. = 5.30e-03
s.d. = 9.80e-03
(b) Box plot for CC metric.
HammerDrive RMDN
Method
0.46
0.48
0.50
0.52
0.54
0.56
0.58
0.60
Sim
s.d. = 2.90e-03
s.d. = 9.00e-03
(c) Box plot for Sim metric.
HammerDrive RMDN
Method
3.2
3.4
3.6
3.8
4.0
4.2
4.4
4.6
4.8
5.0
IG
s.d. = 1.35e-01
s.d. = 2.50e-01
(d) Box plot for IG metric.
Fig. 7. Performance comparison between task-aware HammerDrive and RMDN. Except for
KL
, where lower values correspond to predictions closer to the
ground-truth, for all other parameters higher values correspond to better performance, as presented in detail in Section
VII-A
. For each box, the central line
represents the median, the edges of the box correspond to the 25th and 75th percentiles and the whiskers identify minimum and maximum values.
(a) Changing lane to the left. (b) Lane maintenance.
(c) Changing lane to the right. (d) Failure case.
Fig. 8. Qualitative performance of HammerDrive during the three driving tasks. For each maneuver, the image on the left depicts the scene observed by the
driver. The right plot shows the sequence (10 steps) of ground-truth gaze locations (red dots) and the corresponding sequence of predicted visual attention
distributions P(t), together with their highest mode (blue dots).
been computed, this information is provided to ParRMDN
which infers where drivers should be focusing their attention
to safely complete such a maneuver. These assumptions lead to
a reactive implementation of driver visual attention prediction.
Although the use of telemetry-based maneuver tracking
for task-awareness has shown impressive results in the lit-
erature [
20
], numerous studies have proven that there is a
strong correlation between driver gaze patterns and future ma-
neuvers [
58
], [
59
]. In fact, a lane change is often characterized
by glances to interior and side mirror and to the side window
before any signs of the maneuver appear in the telemetry data.
Therefore, the inclusion of gaze information as an additional
input to HammerDrive opens to interesting directions for future
studies, as it could potentially lead to better and more diverse
maneuver-tracking. Additionally, providing the system with
information on gaze history and driver future intentions could
be leveraged to perform better visual focus predictions.
For our study, we collected the gaze patterns of multiple
drivers in a high-speed highway driving simulator and without
external distractions. In this scenario, we assumed that the
driving task could be divided into a set of simple maneuvers
and implemented HAMMER, a network for real-time maneuver
tracking. Our results showed that HAMMER is a flexible
network, allowing for both kinematic-based models and neural
network-based components, and that it can reliably track lane
change maneuvers. Although the application of HammerDrive
to lateral control maneuvers is supported by previous studies
from the literature [
19
], [
58
], it would be worth studying how
HAMMER can be expanded to track a larger set of maneuvers,
such as turns and accelerations. In these scenarios, drivers
have complete control over the speed of the vehicle, directly
impacting both the duration and the diversity of maneuvers.
These considerations go beyond expanding the maneuver
tracking capabilities of HAMMER. Maneuvers for longitudinal
control of the vehicle, such as braking and coasting, lead to
more diverse, yet equally correct, set of scan patterns. Also,
while HammerDrive can predict multimodal gaze distributions,
our results showed that driver gaze patterns tended to be
predominantly unimodal in the proposed scenario, due to the
presence of single obstacles ahead of the driver. Therefore,
AMADORI et al.: HAMMERDRIVE: A TASK-AWARE DRIVING VISUAL ATTENTION MODEL 11
investigating how HammerDrive can adapt and scale to more
complex scenarios represents a clear challenge for the future.
In this paper, we have addressed the problem of driver focus
prediction as a task that jointly involves top-down processes,
i.e., the maneuver-tracking via our HAMMER network, and
bottom-up processes, i.e., the visual-attention modeling via
our ParRMDN. To achieve this, our experiment was designed
to collect the gaze patterns of multiple human drivers, while
they experienced a similar environment and performed the
same maneuvers. Under these assumptions, we have shown
that HammerDrive is able to reliably predict driver gaze and
that task-awareness plays a critical role for this. Despite the
flexibility offered HammerDrive, driver gaze modeling still
represents a very challenging and complex task, as human
drivers are characterized by different gaze patterns, even when
performing the same maneuver. Among these, we identify three
fundamental aspects that future models should consider and are
often not taken into account when modeling visual attention
in driving: self-pacing, spare capacity and peripheral vision.
Drivers can modulate the complexity of the driving task by
self-pacing in order to meet additional task demands, without
leading them into a distracted state [
60
]. In a similar way,
while driving, humans still have additional spare capacity, as
proven by [
61
], where drivers were able to occlude their vision
while driving without impacting their safety. Understanding
and modeling these concepts could have tremendous benefits,
as the ability to know when and where to glance off-road and
correctly anticipate hazardous situations is a key ability for safe
driving [
62
]. Finally, it would also be interesting to evaluate
and model the role of peripheral vision, and its differences with
foveal vision, in driving, as studies have shown that humans
can use peripheral vision to complete numerous tasks [63].
IX. CONCLUSIONS
In this paper, we addressed the problem of task-aware visual
attention prediction in driving. To solve this, we proposed
HammerDrive, a learnable architecture that uses easy-to-access
data from the vehicle. We developed a realistic virtual-reality
driving simulator and collected a dataset of multimodal data
from a cohort of 20 participants. We performed extensive
experiments and compared HammerDrive against a state-of-
the-art deep learning model for visual attention prediction.
Our results indicate that task-awareness is beneficial for visual
attention prediction and that it can be leveraged using telemetry
data to achieve more robust and reliable predictions.
Our study focused on the application of HammerDrive in a
highway driving scenario and without external distractions. The
extension of the simulated scenario with dynamic obstacles
and a more complex environment represents an important
direction for future works. Additionally, it would be interesting
to investigate how HammerDrive is affected when considering
additional factors, such as distraction and different cognitive
states of the driver.
ACK NOW LE DG ME NT S
This work was supported in part by UK DSTL/EPSRC Grant
EP/P008461/1, and a Royal Academy of Engineering Chair in
Emerging Technologies to Yiannis Demiris.
REFERENCES
[1]
National Highway Traffic Safety Administration, Traffic Safety Facts:
Distracted Driving in Fatal Crashes, 2019.
[2]
Y. Liao, G. Li, S. E. Li, B. Cheng, and P. Green, “Understanding driver
response patterns to mental workload increase in typical driving scenarios,
IEEE Access, vol. 6, pp. 35 890–35 900, 2018.
[3]
Y.-K. Wang, T.-P. Jung, and C.-T. Lin, “EEG-based attention tracking
during distracted driving,IEEE Trans. Neural Syst. Rehabil. Eng., vol. 23,
no. 6, pp. 1085–1094, 2015.
[4]
B. Morris, A. Doshi, and M. Trivedi, “Lane change intent prediction for
driver assistance: On-road design and evaluation,” in IEEE Intell. Veh.
Symp., 2011, pp. 895–901.
[5]
A. Palazzi, D. Abati, S. Calderara, F. Solera, and R. Cucchiara,
“Predicting the Driver’s Focus of Attention: the DR(eye)VE Project,
IEEE Trans. Pattern Anal. Mach. Intell., vol. 41, no. 7, pp. 1720–1733,
2018.
[6]
T. Deng, K. Yang, Y. Li, and H. Yan, “Where does the driver look?
top-down-based saliency detection in a traffic driving environment,IEEE
Trans. Intell. Transp. Syst., vol. 17, no. 7, pp. 2051–2062, 2016.
[7]
A. Borji and L. Itti, “State-of-the-art in visual attention modeling,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 35, no. 1, pp. 185–207, 2013.
[8]
L. Itti, C. Koch, and E. Niebur, “A model of saliency-based visual
attention for rapid scene analysis,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 20, no. 11, pp. 1254–1259, 1998.
[9]
C. Wloka, I. Kotseruba, and J. K. Tsotsos, “Active fixation control to
predict saccade sequences,” in IEEE Conf. Comput. Vis. Pattern Recognit.,
2018, pp. 3184–3193.
[10]
N. D. B. Bruce, C. Catton, and S. Janjic, “A deeper look at saliency:
Feature contrast, semantics, and beyond,” in IEEE Conf. Comput. Vis.
Pattern Recognit., 2016, pp. 516–524.
[11]
A. Borji, “Saliency prediction in the deep learning era: Successes and
limitations,” IEEE Trans. Pattern Anal. Mach. Intell., 2019, to appear.
[12]
J. Coull, “fMRI studies of temporal attention: allocating attention within,
or towards, time,Cogn. Brain Res., vol. 21, no. 2, pp. 216–226, 2004.
[13]
M. Cornia, L. Baraldi, G. Serra, and R. Cucchiara, “Predicting human
eye fixations via an LSTM-based saliency attentive model,IEEE Trans.
Image Process, vol. 27, no. 10, pp. 5142–5154, 2018.
[14]
J. M. Wolfe and T. S. Horowitz, “What attributes guide the deployment
of visual attention and how do they do it?” Nat. Rev. Neurosci., vol. 5,
no. 6, pp. 495–501, 2004.
[15]
H. Hadizadeh and I. V. Bajic, “Saliency-aware video compression,IEEE
Trans. Image Process, vol. 23, no. 1, pp. 19–33, 2014.
[16]
T. Fischer, H. Jin Chang, and Y. Demiris, “RT-GENE: Real-Time Eye
Gaze Estimation in Natural Environments,” in Eur. Conf. Comput. Vis.,
2018, pp. 339–357.
[17]
H. Admoni and B. Scassellati, “Social eye gaze in human-robot
interaction: A review,” J. of Human-Robot Interact., vol. 6, no. 1, pp.
25–63, 2017.
[18]
K. Bengler, K. Dietmayer, B. Farber, M. Maurer, C. Stiller, and
H. Winner, “Three decades of driver assistance systems: Review and
future perspectives,IEEE Intell. Transp. Syst. Mag., vol. 6, no. 4, pp.
6–22, 2014.
[19]
J. Nilsson, J. Silvlin, M. Brannstrom, E. Coelingh, and J. Fredriksson,
“If, when, and how to perform lane change maneuvers on highways,
IEEE Intell. Transp. Syst. Mag., vol. 8, no. 4, pp. 68–78, 2016.
[20]
Y. Xing, C. Lv, H. Wang, H. Wang, Y. Ai, D. Cao, E. Velenis, and
F. Wang, “Driver lane change intention inference for intelligent vehicles:
Framework, survey, and challenges,IEEE Trans. on Veh. Tech., vol. 68,
no. 5, pp. 4377–4390, 2019.
[21]
B. Fr
¨
ohlich, M. Enzweiler, and U. Franke, “Will this car change the
lane? Turn signal recognition in the frequency domain,” in IEEE Intell.
Veh. Symp., 2014, pp. 37–42.
[22]
A. Jain, H. S. Koppula, B. Raghavan, S. Soh, and A. Saxena, “Car that
knows before you do: Anticipating maneuvers via learning temporal
driving models,” in IEEE Int. Conf. Comput. Vis., 2015, pp. 3182–3190.
[23]
P. Kumar, M. Perrollaz, S. Lefevre, and C. Laugier, “Learning-based
approach for online lane change intention prediction,” in IEEE Intell.
Veh. Symp., 2013, pp. 797–802.
[24]
Y. Liu, P. Lasang, S. Pranata, S. Shen, and W. Zhang, “Driver pose
estimation using recurrent lightweight network and virtual data augmented
transfer learning,” IEEE Trans. Intell. Transp. Syst., vol. 20, no. 10, pp.
3818–3831, 2019.
[25]
Y. Demiris and B. Khadhouri, “Hierarchical attentive multiple models
for execution and recognition of actions,Robot. Auton. Syst., vol. 54,
no. 5, pp. 361–369, 2006.
12 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
[26]
L. Bazzani, L. Torresani, and H. Larochelle, “Recurrent mixture density
network for spatiotemporal visual attention,” in Int. Conf. Learn. Repres.,
2017.
[27]
F. Vicente, Z. Huang, X. Xiong, F. De la Torre, W. Zhang, and D. Levi,
“Driver gaze tracking and eyes off the road detection system,IEEE
Trans. Intell. Transp. Syst., vol. 16, no. 4, pp. 2014–2027, 2015.
[28]
T. Billah, S. M. Rahman, M. O. Ahmad, and M. Swamy, “Recognizing
distractions for assistive driving by tracking body parts,IEEE Trans.
Circuits and Syst. for Video Tech., vol. 29, no. 4, pp. 1048–1062, 2018.
[29]
Y. Xing, C. Lv, H. Wang, D. Cao, E. Velenis, and F.-Y. Wang, “Driver
activity recognition for intelligent vehicles: A deep learning approach,
IEEE Trans. Veh. Tech., vol. 68, no. 6, pp. 5379–5390, 2019.
[30]
Y. Xia, D. Zhang, J. Kim, K. Nakayama, K. Zipser, and D. Whitney,
“Predicting driver attention in critical situations,” in Asian Conf. Comput.
Vis., 2018, pp. 658–674.
[31]
T. Deng, H. Yan, L. Qin, T. Ngo, and B. S. Manjunath, “How do
drivers allocate their potential attention? Driving fixation prediction via
convolutional neural networks,IEEE Trans. Intell. Transp. Syst., 2019.
[32]
T. Deng, H. Yan, and Y.-J. Li, “Learning to boost bottom-up fixation
prediction in driving environments via random forest,IEEE Trans. Intell.
Transp. Syst., vol. 19, no. 9, pp. 3059–3067, 2017.
[33]
A. Tawari and B. Kang, “A computational framework for driver’s visual
attention using a fully convolutional architecture,” in IEEE Intell. Veh.
Symp., 2017, pp. 887–894.
[34]
Y. Yao, M. Xu, Y. Wang, D. J. Crandall, and E. M. Atkins, “Unsupervised
traffic accident detection in first-person videos,” in IEEE Intern. Conf.
on Intell. Robots and Syst. IEEE, 2019, pp. 273–280.
[35]
M. Land and B. Tatler, Looking and acting: Vision and eye movements
in natural behaviour. Oxford University Press, 2009.
[36]
M. M. Hayhoe, “Vision and action,Annu. Rev. Vis. Sci., vol. 3, pp.
389–413, 2017.
[37]
M. Da Lio, F. Biral, E. Bertolazzi, M. Galvani, P. Bosetti, D. Windridge,
A. Saroldi, and F. Tango, “Artificial Co-Drivers as a universal enabling
technology for future intelligent vehicles and transportation systems,”
IEEE Trans. Intell. Transp. Syst., vol. 16, no. 1, pp. 244–263, 2015.
[38]
M. Da Lio, A. Mazzalai, and M. Darin, “Cooperative intersection support
system based on mirroring mechanisms enacted by bio-inspired layered
control architecture,” IEEE Trans. Intell. Transp. Syst., vol. 19, no. 5,
pp. 1415–1429, 2018.
[39]
M. Da Lio, A. Mazzalai, K. Gurney, and A. Saroldi, “Biologically guided
driver modeling: The stop behavior of human car drivers,” IEEE Trans.
Intell. Transp. Syst., vol. 19, no. 8, pp. 2454–2469, 2018.
[40]
C. Sentouh, A. Nguyen, M. A. Benloucif, and J. Popieul, “Driver-
Automation cooperation oriented approach for shared control of lane
keeping assist systems,” IEEE Trans. Control Syst. Technol., vol. 27,
no. 5, pp. 1962–1978, 2019.
[41]
D. D. Salvucci, “Modeling driver behavior in a cognitive architecture,”
Human Factors, vol. 48, no. 2, pp. 362–380, 2006.
[42]
D. D. Salvucci and R. Gray, “A two-point visual control model of
steering,” Perception, vol. 33, no. 10, pp. 1233–1248, 2004.
[43]
J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C. Lebiere, and
Y. Qin, “An integrated theory of the mind.Psychological review, vol.
111, no. 4, p. 1036, 2004.
[44]
P. Bosetti, M. Da Lio, and A. Saroldi, “On curve negotiation: From
driver support to automation,IEEE Trans. Intell. Transp. Syst., vol. 16,
no. 4, pp. 2082–2093, 2015.
[45]
D. Shin, B. Kim, K. Yi, A. Carvalho, and F. Borrelli, “Human-centered
risk assessment of an automated vehicle using vehicular wireless
communication,” IEEE Trans. Intell. Transp. Syst., vol. 20, no. 2, pp.
667–681, 2019.
[46]
D. Ognibene and Y. Demiris, “Towards active event recognition.” in Int.
Joint Conf. on Artif. Intell., 2013.
[47]
A. Graves, “Generating sequences with recurrent neural networks,
arXiv:1308.0850, pp. 1–43, 2013.
[48]
D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learning
spatiotemporal features with 3D convolutional networks,” in IEEE Int.
Conf. Comput. Vis., 2015, pp. 4489–4497.
[49]
M. Kok, J. Hol, and T. Sch
¨
on, “Using inertial sensors for position and
orientation estimation,” Found. and Trends in Signal Process., vol. 11,
pp. 1–153, 2017.
[50]
J. Kong, M. Pfeiffer, G. Schildbach, and F. Borrelli, “Kinematic and
dynamic vehicle models for autonomous driving control design,” in IEEE
Intell. Veh. Symp., 2015, pp. 1094–1099.
[51]
S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
[52]
A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V. Koltun, “Carla:
An open urban driving simulator,” in Conf. on Robot Learn., 2017.
[53]
S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual
and physical simulation for autonomous vehicles,” in Field and Service
Robotics, 2018, pp. 621–635.
[54]
N. Pugeault and R. Bowden, “How much of driving is preattentive?”
IEEE Trans. Veh. Technol., vol. 64, no. 12, pp. 5424–5438, 2015.
[55]
W. Wang, J. Shen, J. Xie, M. Cheng, H. Ling, and A. Borji, “Revisiting
video saliency prediction in the deep learning era,” IEEE Trans. Pattern
Anal. and Mach. Intell., 2019, to appear.
[56]
S. Gorji and J. J. Clark, “Going from image to video saliency: Augmenting
image salience with dynamic attentional push,” in IEEE Conf. on Comput.
Vis. and Pattern Recog., 2018, pp. 7501–7511.
[57]
M. Kummerer, T. S. Wallis, and M. Bethge, “Saliency benchmarking
made easy: Separating models, maps and metrics,” in Eur. Conf. Comput.
Vis., 2018, pp. 770–787.
[58]
S. Martin, S. Vora, K. Yuen, and M. M. Trivedi, “Dynamics of driver’s
gaze: Explorations in behavior modeling and maneuver prediction,IEEE
Trans. Intell. Veh., vol. 3, no. 2, pp. 141–150, 2018.
[59]
A. Doshi and M. M. Trivedi, “On the roles of eye gaze and head dynamics
in predicting driver’s intent to change lanes,” IEEE Trans. Intell. Transp.
Syst., vol. 10, no. 3, pp. 453–462, 2009.
[60]
K. Kircher and C. Ahlstrom, “Minimum required attention: A human-
centered approach to driver inattention,Hum. Factors, vol. 59, no. 3,
pp. 471–484, 2017.
[61]
K. Kircher, T. Kujala, and C. Ahlstr
¨
om, “On the difference between
necessary and unnecessary glances away from the forward roadway: An
occlusion study on the motorway,Hum. Factors, vol. 62, no. 7, pp.
1117–1131, 2020.
[62]
C. C. McDonald, A. H. Goodwin, A. K. Pradhan, M. R. Romoser, and
A. F. Williams, “A review of hazard anticipation training programs for
young drivers,J. of Adolesc. Health, vol. 57, no. 1, pp. S15–S23, 2015.
[63]
R. Rosenholtz, “Capabilities and limitations of peripheral vision,” Annu.
Rev. Vis. Sci., 2016.
Pierluigi Vito Amadori
(S’14, M’17) received the
M.Sc. degree (Hons.) in Telecommunications Engi-
neering from the University of Rome La Sapienza,
Rome, Italy, in 2013 and the Ph.D. degree in
Electronic Engineering from the Department of Elec-
trical & Electronic Engineering, University College
London, London, U.K., in 2017.
He currently holds a position as a Postdoctoral
Research Associate at the Personal Robotics Lab-
oratory at Imperial College London, London, U.K.
His main research interests include driver monitoring,
user modeling and driving assistance systems.
Tobias Fischer
(M’16) received the B.Sc. degree
from Ilmenau University of Technology, Germany,
in 2013, the M.Sc. degree in Artificial Intelligence
from the University of Edinburgh, U.K., in 2014, and
the Ph.D. degree from the Personal Robotics Lab,
Imperial College London, London, U.K, in 2018.
His research interests include both computer vision
and human vision, visual attention and computational
cognition. He is interested in applying this knowledge
to cognitive robotics.
Dr. Fischer was a recipient of the Queen Mary
Award for the Best U.K. Robotics PhD Thesis in 2018 and the Eryl Cadwaladr
Davies prize for the best departmental thesis in 2017-2018.
Yiannis Demiris
(SM’03) received the B.Sc. (Hons.)
degree in artificial intelligence and computer science
and the Ph.D. degree in intelligent robotics from
the Department of Artificial Intelligence, University
of Edinburgh, Edinburgh, U.K., in 1994 and 1999,
respectively.
He is a Professor with the Department of Electrical
and Electronic Engineering, Imperial College London,
London, U.K., where he is the Royal Academy of
Engineering Chair in Emerging Technologies, and
the Head of the Personal Robotics Laboratory. His
current research interests include human-robot interaction, machine learning,
user modeling, and assistive robotics. He has published more than 200 journal
and peer-reviewed conference papers in the above areas.
Prof. Demiris was a recipient of the Rectors Award for Teaching Excellence
in 2012 and the FoE Award for Excellence in Engineering Education in 2012.
He is a Fellow of IET, BCS, and Royal Statistical Society.
... For example, [188], [189] use vanishing point because drivers often focus on it to get an optimal view of the road ahead [25], [193]. Other options include drivers' actions [190]- [192], current driving task [194], [195], previous fixation locations [190]- [192], and vehicle telemetry [190]- [192], [196]. Other common features include semantic segmentation maps [197]- [199] and detected objects [196], [200]- [202]. ...
... For instance, optical flow is useful for identifying the direction and magnitude of motion in the scene [198], [203]- [205]. More recently, following successful applications in video action recognition problems [206], 3D convolutional networks have become a popular choice for encoding spatiotemporal data [194], [199], [202], [207]. Some approaches use recurrent networks, combined with individually encoded frames [208], [209] or with a set of frames processed via 3D convolutional layers [197]. ...
... Two recent models, HammerDrive [194] and MEDIRL [195], demonstrate that explicitly modeling the underlying driving task is beneficial for gaze prediction performance. In HammerDrive, a separate module recognizes maneuvers (lane change and lane-keeping) from the vehicle telemetry. ...
Full-text available
Article
Driving safety has been a concern since the first cars appeared on the streets. Driver inattention has been singled out as a major cause of accidents early on. This is hardly surprising, as drivers routinely perform other tasks in addition to controlling the vehicle. Decades of research into what causes lapses or misdirection of drivers’ attention resulted in improvements in road safety through better design of infrastructure, driver training programs, in-vehicle interfaces, and, more recently, the development of driving assistance systems (ADAS) and driving automation. This review focuses on the methods for modeling and detecting spatio-temporal aspects of drivers’ attention, i.e. where and when they look, for the two latter categories of applications. We start with a brief theoretical background on human visual attention, methods for recording and measuring attention in the driving context, types of driver inattention, and factors causing it. We then discuss machine learning approaches for 1) modeling gaze for assistive and self-driving applications and 2) detecting gaze for driver monitoring. Following the overview of state-of-the-art models, we provide an extensive list of publicly available datasets that feature recordings of drivers’ gaze and other attention-related annotations. We conclude with a general overview of the remaining challenges, such as data availability and quality, evaluation methods, and the limited scope of attention modeling, and outline steps toward rectifying some of these issues. Categorized and annotated lists of the reviewed models and datasets are available at https://github.com/ykotseruba/attention_and_driving
... Palazzi's uses C3D network [17,4]. The model of [18] also uses C3D convolution operation. ...
Full-text available
Preprint
The problem of predicting driver attention from the driving perspective is gaining the increasing research focuses due to its remarkable significance for autonomous driving and assisted driving systems. Driving experience is extremely important for driver attention prediction, a skilled driver is able to effortlessly predict oncoming danger (before it becomes salient) based on driving experience and quickly pay attention on the corresponding zones. However, the nonobjective driving experience is difficult to model, so a mechanism simulating driver experience accumulation procedure is absent in existing methods, and the existing methods usually follow the technique line of saliency prediction methods to predict driver attention. In this paper, we propose a FeedBack Loop Network (FBLNet), which attempts to model the driving experience accumulation procedure. By over-and-over iterations, FBLNet generates the incremental knowledge that carries rich historically-accumulative long-term temporal information. The incremental knowledge to our model is like the driving experience to humans. Under the guidance of the incremental knowledge, our model fuses the CNN feature and Transformer feature that are extracted from the input image to predict driver attention. Our model exhibits solid advantage over existing methods, achieving an average 10.3% performance improvement on three public datasets.
... They found that fixations are geared towards obtaining the most valuable task-specific information and thus the subject may not necessarily fixate on the most salient regions in a scene [10,11]. The effect of task on visual attention has given rise to task-driven attention models, which have shown to be effective for visual attention prediction [2,8]. ...
Full-text available
Preprint
The task of driver attention prediction has drawn considerable interest among researchers in robotics and the autonomous vehicle industry. Driver attention prediction can play an instrumental role in mitigating and preventing high-risk events, like collisions and casualties. However, existing driver attention prediction models neglect the distraction state and intention of the driver, which can significantly influence how they observe their surroundings. To address these issues, we present a new driver attention dataset, CoCAtt (Cognitive-Conditioned Attention). Unlike previous driver attention datasets, CoCAtt includes per-frame annotations that describe the distraction state and intention of the driver. In addition, the attention data in our dataset is captured in both manual and autopilot modes using eye-tracking devices of different resolutions. Our results demonstrate that incorporating the above two driver states into attention modeling can improve the performance of driver attention prediction. To the best of our knowledge, this work is the first to provide autopilot attention data. Furthermore, CoCAtt is currently the largest and the most diverse driver attention dataset in terms of autonomy levels, eye tracker resolutions, and driving scenarios.
Article
Driver attention modeling is a crucial technique in building human-centric intelligent driving systems. Considering the human visual mechanism, this study leverages multi-level visual content, including low-level texture features, middle-level optical flows, and high-level semantic information, as the model input. Subsequently, a heterogeneous model is proposed to handle the multi-level input, which integrates the graph and convolutional neural networks. Distinguished from the existing studies that use semantic segmentation, our study directly leverages the objection detection information in an interpretable manner. To deal with the detected objects, in this work, a graph attention network is used to explicitly construct the semantic information, rather than handle the features extracted by convolutional modules for building the latent space features, which are used in existing studies. Further, a semantic attention module is proposed to integrate the non-Euclidean output of the graph network with the Euclidean feature maps of the convolutional neural networks. Finally, these integrated features are decoded to generate a driver attention map. Three typical datasets are used to validate the proposed method. A comprehensive comparison and analysis have proven the feasibility and validity of our proposed method, as well as its ability to achieve state-of-the-art performance.
Full-text available
Article
The aim of this study is to identify the best practices and future research directions for driver lane-change intention (DLCI) prediction using eye-tracking technologies based on a systematic literature review. We searched five academic literature databases and then conducted an in-depth review, structured coding, and analysis of 40 relevant articles. The literature on DLCI prediction is summarized in terms of input features, feature extraction and prediction time windows, labeling methods, and machine learning algorithms. The results show that eye tracking data features along with other data sources can be useful inputs for the prediction of DLCI. Major challenges in this line of research include determining the optimal time window for feature extraction and developing and evaluating the appropriate machine learning algorithm. Suggestions for future research and practice for DLCI prediction in intelligent vehicles are discussed.
Article
Driving is a highly dynamic task requiring extensive and fast processing of visuospatial information in brief time intervals, while experienced drivers can efficiently perform this task. Therefore, investigating the human-like visual fixation behavior shows much potential in developing the advanced driving assistance systems and efficient autonomous vehicles. Literature shows that driver’s visual fixation behavior is guided by a combination of top-down and bottom-up mechanisms. Currently, existing excellent computational frameworks rely heavily on the deep neural network’s strong feature extraction ability, which is inherently task-unaware. These black-box models are limited in strengthening the guidance of task-aware top-down factors in governing the fixation behavior. Furthermore, using the black-box model to fit the training data directly decreases its interpretability. In this paper, we encourage to adopt the white-box method. By researching the findings in the intersection of neuroscience and psychology, we propose to identify two white-box representations for driving separately: the information selection representation including salience, effort, expectancy, and value, and the driving task-aware representation highlighting sub-task and speed. We present a novel computational framework to incorporate all the above factors to predict driver’s visual fixation systematically. On the BDDA dataset, we have conducted the extensive quantitative evaluation and qualitative analysis to validate the effectiveness. Its performance surpasses the previous state-of-the-art by +11.3% in the Pearson’s Correlation Coefficient, especially by a large margin of +24.5% and +28.8% in the turning operations. The ablation study further demonstrates the influence of each component in our framework. Moreover, the trained model has been deployed on the DR(eye)VE dataset to verify the generalization ability, demonstrating the potential of domain adaptation.
Full-text available
Article
Predicting where people look in static scenes, a.k.a visual saliency, has received significant research interests recently. However, relatively less effort has been spent in understanding visual attention over dynamic scenes. This work makes three contributions to video saliency research. First, we introduce a new benchmark, called DHF1K, for predicting fixations during dynamic scene free-viewing, which is a long-time need in this field. DHF1K consists of 1K high-quality, elaborately selected videos annotated by 17 observers using an eye tracker. The videos span a wide range of scenes, motions, object types and backgrounds. Second, we propose a novel video saliency model, called ACLNet, that augments the CNN-LSTM network with a supervised attention mechanism to enable fast end-to-end learning. The attention mechanism explicitly encodes static saliency information, thus allowing LSTM to focus on learning a more flexible temporal saliency representation. Such a design leverages existing large-scale static fixation datasets, avoids overfitting, and significantly improves training efficiency and testing performance. Third, we perform an extensive evaluation of state-of-the-art saliency models on three current datasets (i.e., DHF1K, Hollywood2, UCF sports). Experimental results over more than 1.2K testing videos containing 400K frames demonstrate that ACLNet outperforms other contenders and has a fast processing speed (40fps).
Full-text available
Article
The traffic driving environment is a complex and dynamic changing scene in which drivers have to pay close attention to salient and important targets or regions for safe driving. Modeling drivers' eye movements and attention allocation in traffic driving can also help guiding unmanned intelligent vehicles. However, until now, few studies have modeled drivers' true fixations and allocations while driving. To this end, we collect an eye tracking dataset from a total of 28 experienced drivers viewing 16 traffic driving videos. Based on the multiple drivers' attention allocation dataset, we propose a convolutional-deconvolutional neural network (CDNN) to predict the drivers' eye fixations. The experimental results indicate that the proposed CDNN outperforms the state-of-the-art saliency models and predicts drivers' attentional locations more accurately. The proposed CDNN can predict the major fixation location and shows excellent detection of secondary important information or regions that cannot be ignored during driving if they exist. Compared with the present object detection models in autonomous and assisted driving systems, our human-like driving model does not detect all of the objects appearing in the driving scenes, but it provides the most relevant regions or targets, which can largely reduce the interference of irrelevant scene information.
Full-text available
Chapter
Robust driver attention prediction for critical situations is a challenging computer vision problem, yet essential for autonomous driving. Because critical driving moments are so rare, collecting enough data for these situations is difficult with the conventional in-car data collection protocol—tracking eye movements during driving. Here, we first propose a new in-lab driver attention collection protocol and introduce a new driver attention dataset, Berkeley DeepDrive Attention (BDD-A) dataset, which is built upon braking event videos selected from a large-scale, crowd-sourced driving video dataset. We further propose Human Weighted Sampling (HWS) method, which uses human gaze behavior to identify crucial frames of a driving dataset and weights them heavily during model training. With our dataset and HWS, we built a driver attention prediction model that outperforms the state-of-the-art and demonstrates sophisticated behaviors, like attending to crossing pedestrians but not giving false alarms to pedestrians safely walking on the sidewalk. Its prediction results are nearly indistinguishable from ground-truth to humans. Although only being trained with our in-lab attention data, the model also predicts in-car driver attention data of routine driving with state-of-the-art accuracy. This result not only demonstrates the performance of our model but also proves the validity and usefulness of our dataset and data collection protocol.
Article
Visual saliency models have enjoyed a big leap in performance in recent years, thanks to advances in deep learning and large scale annotated data. Despite enormous effort and huge breakthroughs, however, models still fall short in reaching human-level accuracy. In this work, I explore the landscape of the field emphasizing on new deep saliency models, benchmarks, and datasets. A large number of image and video saliency models are reviewed and compared over two image benchmarks and two large scale video datasets. Further, I identify factors that contribute to the gap between models and humans and discuss remaining issues that need to be addressed to build the next generation of more powerful saliency models. Some specific questions that are addressed include: in what ways current models fail, how to remedy them, what can be learned from cognitive studies of attention, how explicit saliency judgments relate to fixations, how to conduct fair model comparison, and what are the emerging applications of saliency models.
Article
Objective The present study strove to distinguish traffic-related glances away from the forward roadway from non-traffic-related glances while assessing the minimum amount of visual information intake necessary for safe driving in particular scenarios. Background Published gaze-based distraction detection algorithms and guidelines for distraction prevention essentially measure the time spent looking away from the forward roadway, without incorporating situation-based attentional requirements. Incorporating situation-based attentional requirements would entail an approach that not only considers the time spent looking elsewhere but also checks whether all necessary information has been sampled. Method We assess the visual sampling requirements for the forward view based on 25 experienced drivers’ self-paced visual occlusion in real motorway traffic, dependent on a combination of situational factors, and compare these with their corresponding glance behavior in baseline driving. Results Occlusion durations were on average 3 times longer than glances away from the forward roadway, and they varied substantially depending on particular maneuvers and on the proximity of other traffic, showing that interactions with nearby traffic increase perceived uncertainty. The frequency of glances away from the forward roadway was relatively stable across proximity levels and maneuvers, being very similar to what has been found in naturalistic driving. Conclusion Glances away from the forward roadway proved qualitatively different from occlusions in both their duration and when they occur. Our findings indicate that glancing away from the forward roadway for driving purposes is not the same as glancing away for other purposes, and that neither is necessarily equivalent to distraction.
Article
Driver poses recognition contains three tasks such as body joint, head angle, and face landmark estimation, which is of paramount interest for the advanced driver assistance systems (ADAS). Recently proposed methods intend to use deeper and more complicated networks to achieve better performance, which leads to heavy models that are not feasible for the resource limited applications such as ADAS. To resolve this issue, we have worked on the following aspects: 1) a lightweight network model, which is referred to as recurrent multi-task thin net (RM-ThinNet), has been proposed which was especially designed for the computationally and memory limited devices; 2) a recurrent structure has been introduced to handle the scale difference and dependency between different tasks, and this recurrence ensures the different tasks are accomplished at different stages and their outputs can augment each other; and 3) a virtual data synthesization pipeline and a couple transfer learning method have been presented, by which network can be learnt effectively by relatively a small number of real data. Comparisons with the state-of-the-art methods reveal the superiority of the proposed method and competitive performance can be achieved with smaller model size and faster speed.
Article
Intelligent vehicles and advanced driver assistance systems (ADAS) need to have proper awareness of the traffic context, as well as the driver status since ADAS share the vehicle control authorities with the human driver. This paper provides an overview of the ego-vehicle driver intention inference (DII), which mainly focuses on the lane change intention on highways. First, a human intention mechanism is discussed in the beginning to gain an overall understanding of the driver intention. Next, the egovehicle driver intention is classified into different categories based on various criteria. A complete DII system can be separated into different modules,which consist of traffic context awareness, driver states monitoring, and the vehicle dynamic measurement module. The relationship between these modules and the corresponding impacts on the DII are analyzed. Then, the lane change intention inference system is reviewed from the perspective of input signals, algorithms, and evaluation. Finally, future concerns and emerging trends in this area are highlighted.
Article
Intelligent vehicles and advanced driver assistance systems (ADAS) need to have proper awareness of the traffic context as well as the driver status since ADAS share the vehicle control authorities with the human driver. This study provides an overview of the ego-vehicle driver intention inference (DII), which mainly focus on the lane change intention on highways. First, a human intention mechanism is discussed in the beginning to gain an overall understanding of the driver intention. Next, the ego-vehicle driver intention is classified into different categories based on various criteria. A complete DII system can be separated into different modules, which consists of traffic context awareness, driver states monitoring, and the vehicle dynamic measurement module. The relationship between these modules and the corresponding impacts on the DII are analyzed. Then, the lane change intention inference (LCII) system is reviewed from the perspective of input signals, algorithms, and evaluation. Finally, future concerns and emerging trends in this area are highlighted.