Outdoor visual path following experiments.
ABSTRACT In this paper the performance of a topological- metric visual path following framework is investigated in different environments. The framework relies on a monocular camera as the only sensing modality. The path is represented as a series of reference images such that each neighboring pair contains a number of common landmarks. Local 3D geometries are reconstructed between the neighboring reference images in order to achieve fast feature prediction which allows the recovery from tracking failures. During navigation the robot is controlled using image-based visual servoing. The experiments show that the framework is robust against moving objects and moderate illumination changes. It is also shown that the system is capable of on-line path learning.
- [Show abstract] [Hide abstract]
ABSTRACT: This study was planned in order to investigate the effect of left bundle branch block (LBBB) on myocardial velocities obtained by tissue Doppler echocardiography (TDE) and myocardial performance index (MPI). Subjects with LBBB (n = 61) and age-matched healthy subjects (n = 60) were enrolled in the study. Left ventricular (LV) ejection fraction (EF), mitral inflow velocities (E-wave and A-wave), isovolumetric contraction and relaxation time (ICT and IRT), ejection time (ET), and flow propagation velocity (Vp) were measured by conventional echocardiography. Systolic velocity (Sm), early and late diastolic velocities (Em and Am) and time intervals were measured by TDE. MPI was calculated by the formula (ICT + IRT)/ET. LVEF and mitral E/A ratio were similar in both groups. Vp was lower in the LBBB group than in the control group, whereas the E/Em and the E/Vp ratio was higher. LV Sm and Em/Am ratio were lower in LBBB group. Right ventricular Sm and Em/Am ratio were similar in both groups. LV mean and RV MPI were significantly increased in LBBB group. These findings obtained by TDE show that isolated LBBB impairs the ventricular functions. Both of the LV and RV dysfunctions shown by the new parameters may contribute to increased morbidity and mortality in cases with isolated LBBB.Echocardiography 04/2008; 25(3):256-63. · 1.26 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: This paper presents a method for precise 3D en-vironment mapping. It employs only a 3D Time-of-Flight (ToF) camera and no additional sensors. The camera pose is estimated using visual odometry. Imprecision of depth measurements caused by external interfering factors, e.g. sunlight or reflectiv-ities are properly handled by several filters. Pose tracking and mapping is performed on-the-fly during exploration and allows even for hand-guided operation. The final refinement step, comprising error distribution after loop-closure and surface smoothing, further increases the precision of the resulting 3D map 1 .
- [Show abstract] [Hide abstract]
ABSTRACT: This paper describes a system built to enable long-range rover autonomy using a stereo camera as the only sensor. During a learning phase, the system builds a manifold map of overlapping submaps as it is piloted along a route. The map is then used for localization as the rover repeats the route autonomously. The use of local submaps allows the rover to faithfully repeat long routes without the need for an accurate global reconstruction. Path following over nonplanar terrain is handled by performing localization in three dimensions and then projecting this down to a local ground plane associated with the current submap to perform path tracking. We have tested this system in an urban area and in a planetary analog setting in the Canadian High Arctic. More than 32 km was covered—99.6% autonomously—with autonomous runs ranging from 45 m to 3.2 km, all without the use of the global positioning system (GPS). Because it enables long-range autonomous behavior in a single command cycle, visual teach and repeat is well suited to planetary applications, such as Mars sample return, in which no GPS is available. © 2010 Wiley Periodicals, Inc.Journal of Field Robotics 01/2010; 27:534-560. · 2.15 Impact Factor
Outdoor Visual Path Following Experiments
Albert Diosi, Anthony Remazeilles, Siniˇ saˇSegvi´ c and Franc ¸ois Chaumette
Abstract—In this paper the performance of a topological-
metric visual path following framework is investigated in
different environments. The framework relies on a monocular
camera as the only sensing modality. The path is represented
as a series of reference images such that each neighboring pair
contains a number of common landmarks. Local 3D geometries
are reconstructed between the neighboring reference images
in order to achieve fast feature prediction which allows the
recovery from tracking failures. During navigation the robot is
controlled using image-based visual servoing. The experiments
show that the framework is robust against moving objects and
moderate illumination changes. It is also shown that the system
is capable of on-line path learning.
Intelligent autonomous vehicles have performed amazing
feats outdoors. They have driven thousands of kilometers on
freeways , they have navigated on the surface of Mars 
and they have driven over 200km on a challenging desert
route . However, autonomous navigation outdoors using
one camera and no other sensor still remains an exciting
One of the approaches for autonomous navigation using
monocular vision is visual path following. In visual path
following a path to follow can be represented by a series
of reference images and corresponding robot actions (go
forward, turn left, turn right) as in . There a mobile
robot navigated in indoor corridors by applying template
matching to current and reference images and by using the
stored actions. However, storing the robot actions is not
necessary for navigation. In  a robot navigates a 127m
long path outdoors while saving only a series of images from
a camera with a fish-eye lens. To enable pose-based control
of the robot in a global metric coordinate frame, a precise
3D reconstruction of the camera poses is necessary of the
frequently (approx. every 70cm) saved reference images. In
the 3D reconstruction process applied to feature points of the
reference images, a bundle adjustment is used which results
in a long (1 hour) learning phase unsuitable for on-line use.
The length of the path measured by odometry is used to
correct the scale of the map. After learning the path the robot
can very accurately reproduce the path at 50cm/s velocity.
It turns out that reconstructing the robot’s path, or having
3D information is not necessary. In  a robot navigated
140m outdoors at a speed of 35cm/s with 2D image informa-
tion only. During mapping, image features were tracked and
their image patches together with their x image coordinates
The presented work has been performed within the French national
project Predit Mobivip and project Robea Bodega.
The authors are with IRISA/INRIA Rennes, Campus Beaulieu, 35042
Rennes cedex, France. Email: firstname.lastname@example.org
were saved approx. every 60cm traveled. During navigation,
the robot control was based on simple rules applied to the
tracked feature coordinates in the next reference and current
image. The robot however relied on frequent reference image
switches to recover from occlusions due to moving objects.
A person walking across the camera’s field of view between
two reference image switches would have caused a problem
due to covering up each tracked feature.
The work described in  aimed at indoor navigation, can
deal with occlusion at the price of using 3D information.
A local 3D reconstruction is done between two reference
omnidirectional images. During navigation, tracked features
which have been occluded get reprojected back into the
current image. The recovered pose of the robot is used to
guide the robot towards the target image.
Building an accurate and consistent 3D representation
of the environment can also be done using SLAM. For
example in  a robot mapped a 100m path outdoor using
a monocular camera and odometry. There were only 350
features in the map which in our view approaches the
limit that a simple Kalman filter SLAM implementation can
handle in real time on current PCs. However the simulation
result in  of closing million landmark loops predict that
monocular SLAM will be soon a viable choice for creating
accurate maps with large numbers of landmarks.
In this paper the experimental evaluation of a visual path
following framework is presented. This framework is similar
to  in that only local 3D reconstruction is used and
that occluded features get reprojected back into the image.
However the rest of the details are different. For example
in this paper a standard camera is used, tracking is used for
mapping instead of matching, experiments are done outdoors
and the centroids of image features are used to control the
The concept of the framework has been evaluated using
simulations in , while the feature tracker and the com-
plete vision subsystem have been described in , .
As already mentioned, we thus focus in this paper on the
numerous experimental results that have been obtained using
a car-like vehicle.
II. VISUAL NAVIGATION
This section briefly describes the implemented visual
navigation framework. The teaching of the robot i.e. the
mapping of the environment is described first, followed
by the description of the navigation process consisting of
localization and robot control.
sequence of images, i.e. mapping.
The steps involved in building a representation of a path from a
Learning a path (i.e. mapping) starts with the manual
driving of the robot on a reference path while processing
(or storing for off-line mapping) the images from the robot’s
camera. From the images an internal representation of the
path is created, as summarized in fig. 1. The mapping starts
with finding Harris points in the first image, initializing
a Kanade-Lucas-Tomasi (KLT) feature tracker  and by
saving the first image as the first reference image. The KLT1
tracker was modified as proposed in  in order to improve
performance in outdoor sequences acquired from a moving
car. In the tracker postion, scale and contrast parameters of
features are tracked. In the next step a new image is acquired
and the tracked features are updated. The tracking of features
which appear different than in the previous reference image
is abandoned. The rest of the features are then used to
estimate the 3D geometry between the previous reference
and the current image. In the 3D geometry estimation, the
essential matrix is recovered using the calibrated 5 point
algorithm2 used in the MLESAC  random sampling
framework. If the 3D reconstructionerror is low and there are
enough tracked features a new image is acquired. Otherwise
the current image is saved as the next reference image. The
relative pose of the current image with respect to the previous
reference image and the 2D and 3D coordinates of the point
features shared with the previous reference image are also
saved. Then the tracker is reinitialized with new Harris points
added to the old ones and the processing loop continues with
acquiring a new image.
The resulting map (fig. 2) is used during autonomous
navigation in the localization module to provide stable image
1The source code of the KLT tracker maintained by Stan Birchfield can
be found at http://www.ces.clemson.edu/∼stb/klt/
2Free implementation is available in the VW library downloadable from
points for image-based visual servoing.
During navigation, the point features from the map are projected into the
current image and tracked.
The map consists of reference images, 2D and 3D information.
The localization process during navigation is depicted
in fig. 3. The navigation process is started with initial
localization where user selects a reference image close to
the robot’s current location. Then an image is acquired and
matched to the selected reference image. The wide-baseline
matching is done using SIFT descriptors . The estimation
of the camera pose using the matched points enables to
project map points from the reference image into the current
image. The projected points are then used to initialize a KLT
After the initial localization a new image is acquired
and the point positions are updated by the tracker. Using
the tracked points a three-view geometry calculation is
performed between the previous reference, current and next
reference image (fig. 2). If the current image is found to
precede the next reference image, then points from the map
are reprojected into the current image. The projected points
are used to resume the tracking of points currently not
tracked and to stop the tracking of points which are far
from their projections. A new image is acquired next and the
whole cycle continues with tracking. However, if it is found
that the current image comes after the next reference image,
a topological transition is made i.e. the next-next reference
image (fig. 2) becomes the next reference image. The tracker
is then reinitialized with points from the map and the process
continues with acquiring a new image.
Wide-baseline matching is only used outside the initial
localization phase if most features are lost for example due
to a total obstruction of the camera’s field of view. In such
case automatic reinitialization is carried out by matching with
the nearest reference images.
C. Motion Control
In the motion control scheme the robot is not required
to accurately reach each reference image of the path, nor
to follow accurately the learned path since it may not be
useful during navigation. In practice, the exact motion of the
robot should be controlled by an obstacle avoidance module
Localize in map
Fig. 3. Visual localization during navigation.
which we plan to implement soon. Therefore a simple control
algorithm was implemented where the difference in the x-
coordinates (assuming the forward facing camera’s horizon-
tal axis is orthogonal with the axis of robot rotation) of the
centroid of features in the current (xc) and next reference
image (xn) are fed back into the motion controller of the
robot as steering angle Φ:
Φ = − a(xc−xn)
The translational velocity is set to a constant value, except
during turns, where it is reduced (to a smaller constant value)
to ease the tracking of quickly moving features in the image.
Such turns are automatically detected during navigation, by
the analysis of the difference in the feature centroids in the
current, next and next-next image.
Deciding when to stop when reaching the goal position,
is carried out similarly to the reference image switching
strategy of  by observing when the error between current
and last-reference image features starts to rise.
III. EXPERIMENTAL RESULTS
In the experiments a CyCab, a French-made 4 wheel
drive, 4 wheel steered intelligent vehicle designed to carry 2
passengers was used. In our CyCab all computations except
the low-level control were carried out on a laptop with
a 2GHz Centrino processor. A 70◦field of view, forward
looking, B&W Allied Vision Marlin (F-131B) camera was
mounted on the robot at a 65cm height. Except in experiment
3 the camera was used in auto shutter mode, with the rest
of the settings constant.
During all experiments, no software parameters were
changed except that of the forward and turning speed.
Mapping has been performed off-line, except in experiment
6. The image resolution in the experiments was 320x240.
A. Experiment 1
Experiment 1 (see fig. 4) was conducted on an overcast
day with a short time between mapping and navigation. Most
views on the 158m long path contained buildings which
provided stable image features. The main challenges in this
Fig. 4.Paths for experiments 1 and 2.
and numbers) and robot poses (black) during navigation in exp. 1. “R” at
the bottom marks the first reference image pose.
Visual odometry showing 77 reference image poses (green dots
experiment were (i) motion blur in the teaching sequence
caused by fast driving, (ii) driving under a building which
caused a quick illumination change and (iii) people (more
than 10) and cars covering up features during navigation.
In the teaching phase, 958 logged images were reduced
into 77 reference images in 257s (3.7fps). While the robot
was moving at 50cm/s in turns and at 90cm/s otherwise
during navigation, 934 images were processed at 4.1fps on
average. Statistics regarding mapping and navigation are
shown in tab. I. Visual odometry results shown in fig. 5 were
only used for assessing the performance of the system.
The quick illumination change when driving under the
building was easily handled due to the implemented illumina-
tion compensation in the tracker. Motion blur in the teaching
sequence did not impair the performance of the system. The
moving objects and persons did not affect the navigation
because the tracking of features re-appearing after occlusion
were resumed immediately due to the feature reprojection
scheme. Figure 6 contains images processed at the end of
the navigation. They describe an interesting situation where
a moving car progressively occludes most features. It can be
seen that the tracking of re-appearing features is resumed.
B. Experiment 2
Experiment 2 was conducted on a narrow path along a
small lake (fig. 4 and 10). Mapping was carried out in
June, under the strong summer sun. Navigation took place
in October, when vegetation and light conditions were very
different (fig. 8). Despite the large change in the environment,
CyCab managed to navigate about 80% of the path with only
one human intervention.At one place CyCab started brushing
the rose plants on the left side of the path in fig. 8 therefore
we stopped him. Without stopping the vision system, CyCab
Fig. 6. Every second frame of a sequence from experiment 1 demonstrates robust feature (yellow crosses) tracking resumption after occlusion by a passing
the navigation in experiment 2. CyCab completed about 80% of the path.
CyCab could navigate the whole path using a new map (right).
Visual odometry results (left) using a map created 3 month prior
month old reference image (left) and a current image used in navigation in
Large difference in illumination and in the vegetation between a 3
(right) in exp. 2 which the vision system could not handle any more. Notice
the missing flowers in the flowerbed.
Difference between the reference image (left) and current image
Fig. 10.CyCab driving on the narrow path in experiment 2.
was moved 50cm to the right and its automatic motion was
resumed. CyCab’s vision system gave up close to the end
of the track when the change in the environment was too
large (see fig. 9). Even though CyCab did not complete the
whole path (see the left image in fig. 7 where it failed), this
experiment still represents a large success because of the
difficult conditions CyCab could cope.
Shortly after CyCab got lost, we have repeated the exper-
iment using a new map. As it can be seen in the right image
of fig. 7, CyCab completed the path without any problems.
The frame rates during navigation are lower in this exper-
iment (see tab. I), because the vision software was compiled
with the wrong optimization switches and the processor was
running on a lower frequency to save energy.
C. Experiment 3
Fig. 11. The path for experiment 3.
away in experiment 3.
Larger noise in the visual odometry where all features are far
in the current image (right) during navigation in exp. 3.
Sun shining into the camera in the reference image (left), but not
In experiment 3 CyCab completed an approximately 304m
track, where in some places (right side in fig. 11), the closest
features were more than 100m away. The CyCab wide track
enabled us to examine the lateral error in CyCab’s motion
under such conditions. The mapping and navigation part
of the experiment was conducted in succession, under very
bright lighting conditions. Instead of the usual auto-shutter
mode, the camera was used in its high dynamic range mode.
As one can expect, the error in the estimated pose during
navigation was the largest at those places, where there were
no close features. The large pose error is represented by
noisy points in the right bottom part of the path in fig. 12.
The 3D pose error resulted in an early switching of a few
reference images during turning, and subsequently following
the learned path with a 1m lateral error on a short section of
the path. Other than that, CyCab performed excellently even
when the sun was shining into its camera as in fig. 13.
D. Experiment 4
Fig. 14. Visual odometry for the loop closing experiment (exp. 4).
4, but not in the current image (right) during navigation.
Sun shining into the camera in the reference image (left) of exp.
The aim of this experiment was to investigate navigation
in a loop. The teaching was performed by driving CyCab
in a full loop in a circular parking lot of approx. 119m
circumference. The beginning and end of the loop were
closed by matching the first and last image of the teaching
sequence. If neighboring nodes were connected with line
segments, then the first and the last green dot in fig. 14 were
Between mapping and navigation, 4 cars left the parking
lot. One of these cars provided the only close features at the
beginning of the loop, which resulted in noisy pose estimates.
CyCab successfully completed 1.25 loops, while the small
change of illumination (see fig. 15) did not matter.
E. Experiment 5
Fig. 16. The first images during navigation in exp. 5 (left) and in 6 (right).
In exp. 5 the robot drove until the end of the road. In exp. 6 the robot
parked itself into the garage close to the center of the image.
Fig. 17. Visual odometry for experiment 5.
In experiment 5 (see fig. 16 and 17), CyCab completed
a 100m straight path at a fast, 1.8m/s speed. A short video
clip of this experiment (together with the next experiment)
is included as supplementary material.
F. Experiment 6
Fig. 18.Visual odometry for experiment 6.
In this experiment online mapping and a practical appli-
cation is demonstrated. In the current state of the navigation
system, i.e. without obstacle detection-avoidance, etc. the
practical applications are limited. However, even now the
framework can be used for automatic parking in private
properties which are under the control of the user.
During the experiment a map was created on-line while
driving CyCab from the entrance of IRISA to the CyCab
garage approx. 50m away (see fig. 16 and 18) at about
50cm/s. Then CyCab was manually driven to the entrance
of IRISA where the driver got out and CyCab drove itself
into the garage. During mapping clouds covered the sun,
while during navigation the sun was not covered.
By performing simple image-based visual servoing instead
of position-based control of the robot, one can have many
advantages. Since there is no need for an accurate robot pose
during navigation, one can allow a larger 3D reconstruction
error during mapping. Because of this, there is no need
to perform a computationally costly global bundle adjust-
ment and mapping can be performed on-line. During the
experiments it was noticed that, after the baseline between
reference images increased beyond a certain distance, the
3D reconstruction error increased as well. Therefore if a
larger 3D reconstruction error is allowed, then one can have
SUMMARY OF THE VISUAL PATH FOLLOWING EXPERIMENTS
fps v turning
larger distances between reference images, and the memory
requirement for storing the map is reduced. This can be
seen for example in experiment 3 where the average distance
between reference images was 3.1m.
The implemented contrast compensation in the tracker is
able to handle large affine changes of illumination between
the reference and current images which was crucial for
example during experiment 2 (fig. 8).
The use of 3D information enables to resume the tracking
of features just becoming visible after occlusion as can
be seen in fig. 6. This property is important in dynamic
environments. Also, having 3D information also enables to
check the consistency of the tracked features. Tracked points
which “jump” from the background onto a moving object
in the foreground are discarded. Even though having 3D
informationmay not be necessary for path following as stated
in the introduction, it may extend the area of applicability of
an outdoor path following system.
The framework enables the learning and navigation of
long paths since the memory and computational requirements
for mapping grow linearly with the length of the path. The
computational cost during navigation is approx. constant.
The main weakness in the current implementation of the
framework is the reliance on 3D pose to switch reference
images. In cases when there is a large 3D error, it can
happen that a reference image switch is not performed, or
it is performed in the wrong direction. Such misbehavior
occasionally happens when most of the observed points are
located on a plane or on a tree. To address this issue, we are
planning to investigate a reference image switching strategy
based on the more stable image information.
A further limitation is that of the illumination. Extreme
illumination changes such as the sun shining into the camera
during mapping but not during navigation, or the lack of light
may impair the performance of the framework, especially
that of the matcher.
At last, navigation frameworks for uncontrolled environ-
ments such as the one described in this paper should be able
to detect and avoid obstacles. Since this is not implemented
in the framework yet, it constitutes part of the future work.
An experimental evaluation of a framework for visual path
following in outdoor urban environments using only monoc-
ular vision was presented in this paper. In the framework
no other sensor than a camera was used. It was shown that
the use of local 3D information, contrast compensation and
image-based visual servoing can lead to a system capable of
navigating in diverse outdoor environments with reasonably
changing lighting conditions and moving objects.
The help of Fabien Spindler and Andrea Cherubini during
experiments are gratefully acknowledged.
 Z. Chen and S. T. Birchfield. Qualitative vision-based mobile robot
navigation. In ICRA, Orlando, 2006.
 Y. Cheng, M.W. Maimone, and L. Matthies. Visual odometry on the
Mars exploration rovers - a tool to ensure accurate driving and science
imaging. Robotics & Automation Magazine, 13(2), 2006.
 U. Frese and L. Schroder. Closing a million-landmarks loop. In IROS,
 T. Goedeme, T. Tuytelaars, G. Vanacker, M. Nuttin, and L. Van Gool.
Feature based omnidirectional sparse visual path following. In IROS,
Edmonton, Canada, August 2005.
 H. Jin, P. Favaro, and S. Soatto. Real-time feature tracking and outlier
rejection with changes in illumination.
 T Lemaire, C. Berger, I. Jung, and S. Lacroix. Vision-based SLAM:
Stereo and monocular approaches. IJCV/IJRR special joint issue, 2007.
 D. G. Lowe. Distinctive image features from scale-invariant keypoints.
International Journal of Computer Vision, 60(2):91–110, 2004.
 Y. Matsumoto, M. Inaba, and H. Inoue. Visual navigation using view-
sequenced route representation. In ICRA, Minneapolis, April 1996.
 D. Nister. An efficient solution to the five-point relative pose problem.
PAMI, 26(6):756–770, June 2004.
 T. Pomerleau, D.and Jochem. Rapidly adapting machine vision for
automated vehicle steering. IEEE Expert, 11(2), 1996.
 A. Remazeilles, P. Gros, and F. Chaumette. 3D navigation based on
a visual memory. In ICRA’06, 2006.
 E. Royer, J. Bom, M. Dhome, B Thuillot, M. Lhuillier, and F. Mar-
moiton. Outdoor autonomous navigation using monocular vision. In
IROS, pages 3395–3400, Edmonton, Canada, August 2005.
 S. Segvic, A. Remazeilles, and F. Chaumette. Enhancing the point
feature tracker by adaptive modelling of the feature support. In ECCV,
Graz, Austria, 2006.
 S. Segvic, A. Remazeilles, A. Diosi, and F. Chaumette. Large scale
vision based navigation without an accurate global reconstruction. In
Accepted to CVPR’07, Minneapolis, June 2007.
 Jianbo Shi and Carlo Tomasi. Good features to track. In CVPR’94,
pages 593–600, 1994.
 S. Thrun et al. Stanley, the robot that won the DARPA Grand
Challenge. Journal of Field Robotics, 23(9), 2006.
 P. H. S. Torr and A. Zisserman. MLESAC: A new robust estimator
with application to estimating image geometry. Computer Vision and
Image Understanding, 78:138–156, 2000.
In ICCV, volume 1, pages