Content uploaded by Jorge Peña Queralta
Author content
All content in this area was uploaded by Jorge Peña Queralta on Apr 20, 2020
Content may be subject to copyright.
Visual Odometry Offloading in Internet of Vehicles with
Compression at the Edge of the Network
L. Qingqing1,2, J. Pe ˜
na Queralta2, T. N. Gia2, H. Tenhunen3, Z. Zou1and T. Westerlund2
1School of Information Science and Technology, Fudan Universtiy, China
2Department of Future Technologies, University of Turku, Finland
3Department of Electronics, KTH Royal Institute of Technology, Sweden
Emails: 1{qingqingli16, zhuo}@fudan.edu.cn, 1{jopequ, tunggi, tovewe}@utu.fi, 3hannu@kth.se
Abstract—A recent trend in the IoT is to shift from traditional
cloud-centric applications towards more distributed approaches
embracing the fog and edge computing paradigms. In au-
tonomous robots and vehicles, much research has been put into
the potential of offloading computationally intensive tasks to
cloud computing. Visual odometry is a common example, as real-
time analysis of one or multiple video feeds requires significant
on-board computation. If this operations are offloaded, then
the on-board hardware can be simplified, and the battery life
extended. In the case of self-driving cars, efficient offloading can
significantly decrease the price of the hardware. Nonetheless,
offloading to cloud computing compromises the system’s latency
and poses serious reliability issues. Visual odometry offloading
requires streaming of video-feeds in real-time. In a multi-
vehicle scenario, enabling efficient data compression without
compromising performance can help save bandwidth and increase
reliability.
Index Terms—Odometry; VSLAM; Visual Odometry; Visual
SLAM; Internet of Vehicles; IoV; Edge Computing;
I. INTRODUCTION
Accurate localization is one of the key pillars behind full
autonomy. It is also essential for wider types of advanced
intelligent systems, including those related to human-robot
interaction. In terms of self-driving cars, the future of au-
tonomous vehicles is also the future of connected vehicles [1].
This will come under the umbrella of Internet of Everything
(IoE) and, more concretely, the Internet of Vehicles (IoV)
paradigms [2]. In these paradigms, all vehicles share data with
each other in vehicle-to-vehicle (V2V) communication, and
any entity with information that might affect its operation in
vehicle-to-everything (V2X) communication.
In GNSS-denied environments, or in those applications
where high accuracy is necessary, localization often relied on
odometry information. Typical ways of obtaining odometry
information is through lidars or cameras. Visual odometry with
mono or stereo vision has been extensively studied and current
state-of-the-art methods provide robust solutions for accurate
localization in both indoors and outdoors scenarios.
As the first cars with self-driving capabilities are entering
the market, a significant part of the vehicle production cost
is in the hardware required to provide robust and reliable
autonomous operation. V2V and V2X will be one of the key
factors in obtaining the target in terms of reliability, road
safety, traffic efficiency, and energy savings. If we combine
this with intensive computational offloading to near infrastruc-
ture, there is potential for important savings in the on-board
complexity of both hardware and software [3]. This can be
implemented in vehicles with human-in-the-loop, where the
operation can change to manual if required. Nonetheless, in
any case, strict control must be maintained over the network
load in order to ensure that the bandwidth available for each
unit is enough to keep latency and delays within safe limits.
II. REL ATE D WORK
To the best of our knowledge, previous works that have
considered offloading visual odometry calculations have uti-
lized cloud-centric architectures. Yun et al. proposed a
cloud robotics platform named RSE-PF for distributed visual
SLAM [4]. The authors reported round-trip latency of around
150ms. This, compared to state-of-the-art methods able of
processing at 30 frames/second or more, might result in delays
that limit the potential application scenarios. While the authors
utilized websockets in order to save bandwidth compared
to HTTP, they did not report on the maximum number of
concurrent units that could be handled. Dey et al., while
still relying on cloud servers, also proposed offloading in
a multi-tier edge+cloud setup [5]. The authors put a focus
on finding the optimal offloading strategy to make best use
of the different network layers. They formulated an integer
linear programming problem and provided an initial approach
for dynamically deciding on the best offloading decision, in
which the network bandwidth was a variable. In contrast, we
put the focus on studying what is the maximum bandwidth
savings that we can obtain without sacrificing performance,
while maintaining a reliable service with minimal latency.
III. IMAG E COMPRESSION FOR VISUAL ODOMETRY
Visual odometry (VO) is an estimation of camera motion
method based on a series of sequential images. VO can be
applied in a verity applications. The general idea is to calculate
the position correspondences between the two views by finding
some invariant features. In this work we utilize an approach
to visual odometry consisting of 3D-2D correspondences: In
this method, the transformation matrix is calculated using the
Perspective-n-Point(PnP) method. Firstly, the features across
two neighbor frames obtained by the camera are detected
and matched. The best matching points will be obtained
TABLE I
COMPRESSION RATE AND PERFORMANCE IMPACT WITH DIFFERENT JPEG
COMPRESSION TECHNIQUES.
Bandwidth Savings Accuracy loss
JPEG50 22% 0.1%
JPEG10 71% 1.2%
after incorrect matches are discarded. Then 3D points are
obtained by triangulating. After that, we eliminate inaccurate
3D points twice and combined the optical flow method and
feature matching method to find more accurate 3D-to-2D
correspondences. Finally, camera pose will be solved through
these correspondences by PnP algorithm.
A major challenge in offloading visual odometry is the
amount of data that needs to be streamed over the network.
Compared to 2D lidar data or IMU data, a continuous stream
of images consumes significantly higher bandwidth. Therefore,
if images can be compressed without compromising the algo-
rithm’s performance, we can increment the efficiency of the
edge offloading scheme.
IV. EXP ERIMENT AND RESU LTS
We have analyzed how traditional JPEG compression affects
the performance of visual offloading algorithms for different
levels of image quality levels. We have utilized a subset of
the TUM dataset [6]. The experiments have been run with a
set of 3682 acquired over 186 seconds. We compare the errors
produced as a consequence of compressing the images at 10%
quality and 50% quality, which results in compressed sizes
of 0.78 and 0.29 times the original image size, respectively,
as shown in Table 1. In terms of localization accuracy, the
cumulative error is around 10 times smaller in the case of the
50% quality compression. The total and cumulative errors, and
the errors in each direction are shown in Figure 1. We can see
that, in this case, the performance varies across dimensions.
The most clear difference appears in the error in the z axis,
where the difference between the two compression methods
is evident. In general, we can conclude that we can utilize
images with 50% quality without compromising accuracy in
most applications, while a 10% quality can be utilized in
situations where the the accuracy requirements are not so tight.
It is worth mentioning that the drift in the localization error
is continuous in both cases, with a mostly constant average
error. Therefore, it these method should be combined with
loop-closure techniques that ensure that the localization error
can be reduced to near zero within certain intervals of time.
V. CONCLUSION AND FUTURE WORK
In this paper, we present preliminary results on the study of
how different degrees of image compression affect the perfor-
mance of visual odometry algorithms. We have concluded that
around 20% of the network bandwidth can be saved without
compromising accuracy, while a slight reduction in accuracy
can bring over 70% of network load reduction, enabling a
more flexible scaling of the computational offloading scheme.
0 100 200 300
0
2
4
6
·10−3
JPEG10
JPEG50
(a) Error in x axis (m)
0 50 100 150
0
2
4
·10−3
JPEG10
JPEG50
(b) Error in y axis (m)
0 100 200 300
0
2
4
·10−3
JPEG10
JPEG50
(c) Error in z axis (m)
0 100 200 300
0
2
4
6
·10−5
Instant Error Square
JPEG10 Err.
JPEG50 Err.
0 100 200 300
0
1
2
3
·10−2
JPEG10 Cumm. Err.
JPEG50 Cumm. Err.
Fig. 1. Performance comparison of different compression rates.
In future work, we will study a wider range of image
compression techniques, including ML-powered lossless com-
pression techniques and measure network conditions.
ACKNOWLEDGMENT
This work has been supported by NFSC grant No.
61876039, and the Shanghai Platform for Neuromorphic and
AI Chip (NeuHeilium).
REFERENCES
[1] A. Bazzi et al. On the performance of ieee 802.11 p and lte-v2v for
the cooperative awareness of connected vehicles. IEEE Transactions on
Vehicular Technology, 66(11):10419–10432, 2017.
[2] O. Kaiwartya et al. Internet of vehicles: Motivation, layered architecture,
network model, challenges, and future aspects. IEEE Access, 2016.
[3] V. K. Sarker et al. Offloading slam for indoor mobile robots with edge-
fog-cloud computing. In ICASERT, IEEE, 2019.
[4] P. Yun et al. Towards a cloud robotics platform for distributed visual
slam. In Computer Vision Systems. Springer, 2017.
[5] S. Dey et al. Robotic slam: A review from fog computing and mobile
edge computing perspective. In MOBIQUITOUS. ACM, 2016.
[6] J. Sturm et al. Evaluating egomotion and structure-from-motion ap-
proaches using the TUM RGB-D benchmark. In CDCFR, IROS, 2012.
