computer vision for autonomous navigation...computer vision for autonomous navigation june5, 1988...
TRANSCRIPT
COMPUTER VISION FOR AUTONOMOUS NAVIGATION
June5, 1988
Martial Hebert Carnegie-Mellon University
I
Content
1. Perception and mobile robots
2. Sensors
3. Representation of perceptual information for mobile robots
4. Detailed description of an autonomous ve- hicle: the NAVLAB
1
Degrees of Autonomy
Direct Operation
Teleoperation
upervised Operation
Autonomous System
0 0 0 0
0 0 e
2
Applications of autonomous mobile robots (incomplete list, in no particular order)
Indoor navigation (SRI, Stanford, INRIA, LAAS, CMU)
Indoor cleaning (LIFIA, LAAS)
Construction (Rex Excavator (CMU))
On-road navigation (ALV project (CMU, Martin Marietta, Maryland Univ.))
Chauffeur (Munich Univ.)
Undersea exploration (UNH, Texas A&M, NBS (MAW))
Off-road exploration && transportation (Hughes AI, CMU, Martin Marietta (ALV), Ohio State walking machine)
Planetary explorer (CMU, JPL) Surveillance (CMUDenning Mobile Robots)
1
Components of a mobile robot
Navigation Sensed world
3-D world
Predictions Speed Miss ion Steering
World model I Vehicle I control
I Preloaded I knowledge
PERCEPTION FOR NAVIGATION
Find
Track road Find Intersection
MODELING
Observations from multiple sensors and positions
Challenges for perception
0 Uncertainty: Errors may accumulate as the vehicle moves if perception uncertainty is
A A
not explicitly represented.
- Mobility: Variability between vations.
3-D: Obstacles, terrains, ..etc.
sensor obser-
are part of a 3-D world.
1
Sensors
Passive:
-Video camera - Passive stereo vision
Active:
- Sonar -Laser range finder
1
Video cameras
Context: a
a
a
Object tracking: Blob or edge extraction, vehicle tracking &om image to image through feedback to vehicle controller.
Indoor navigation: Edge detection, inter- pretation in highly constraint environment (e.g. Vertical and horizontal edges of walls and furniture).
Road following:
- Road edges extraction and tracking. -Road region extraction for color data.
Main problem: Calibration + A camera provides only 2-D
information whereas the vehicle moves in a 3-D world.
1
Calibration problem
2-D irnaee
Possible solution: Camera model
Flat ground plane Vertical objects
Example: Road following (U. Maryland) I
f
/
Original image (intensity, 256x256)
Edge extraction in small windows (up to 64x64)
Predictions of window locations in image I
Vehcle
Road model . 4 4 m world
1
1 I
Figure 2c. The original image, along with the windows and located road boun- daries
Passive depth estimation
The location of a point can be computed from its projections from differents viewpoints. From different cameras (stereo), or from dif- ferent positions (e.g. motion).
OHOW to
What to
find good matches between images
match ? (features vs. pixels)
Binocular vs, trinocular ?
How to use the vehicle’s motion ?
1
PASSIVE STEREO
Y t z I
Object in world
Direction from left
Direction from right
View from right image
rn la- View from left image - x = f.X/(f - Z) Disparity
1 y = f.Y/(f - Z)
Finding good matches
Similarity measures
Epipolar constraints
Ordering constraints
Coarse-to-fine
Small range of disparity values
0 Overdetermination: trinocular techniques
2
a
a
a
0
a
a
Example: Mobi (Stanford)
Goal: Navigation in hallways
Two-cameras stereo
Features: vertical edges
Initial matches: similarity of grey level curves around edges
Local consistency: similarity of grey level curves between edges
Constraint propagation
i
1
Figrrc 4. Stereo match proposals and grey level curves
X~~nocular stereo
A
f
I
L3 1
A1 and A2 are images of the same point only if there is an A3 at the intersection of L31 and L32.
+ No search.
I
Using vehicle motion
Use depth estimates from previous positions to:
Predict matches and reduce the search
Improve (reduce uncertainty) of current depth estimates
1
f Left image
FIDO (CMU) Right image
Interest points in left image
Prediction boxes for corresponding pomts in nght image
Matching
Computation of 3-D position
Pixel-based (or iconic) techniques
imagek-11 disparity image~l Ccxrcwon Integration
(Example: Matthies' depth from motion)
disparity disparity
variance Regularization ", .ante c
L an >
0 Do not match features
4
predicted '
I I disparity Y
Correlate images directly
I I I Motion I
o f t mdft-1. intensity image
Rediction Varial lCC
d: disparity at pixel
Minimum gives best d and uncertainty VU@). Depth map is updated over time:
Figure 2: ImNc deph estimation block diagram
3
Uncertainty from passive depth estimation
Right camera
a
Sonar
Time-of-return of sound wave. Typical characteristics:
Single depth measurement
Limited to 30 ft.
30° field of view (Polaroid)
Low data rate
Low cost
Applications :
One unit:
-Soft bumper - Surface (e.g. wall) tracking
Sonar ring:
- Obstacle avoidance -Map building (More later on that)
1
Sonar nng (Typically 24 Polaroid units)
Empty region (increased resolution due to overlap)
Emntv
Vehicle
Figure 2: Thc Neptune mobilc robot, with a pair of c a m e m and h e sonar ring. For cxpcri1ncnt.s in scn.wr inicgntion. Ihc c a m c m wwc mounrd lowcr, so Iliai ilicir horizon linc w ~ ~ l l t l hc cltfic to ihc cruss-.~i~onal vicw providd hy ilic sonar ring.
Exceptions
FMC sonar for outdoor vehicle: 30KHz, 64 ft. range, 16 x 24 image
2
Exceptions
Underwater sonar imaging
FIG. 1. Underwater robot EAVE-East (Experimental Autonomous Vehicle) in its launch cradle at the Uni- versity of New Hampshire. The large white cylinders contain electronics: the smaller. lower white cylinders hold batteries: and the darker cylinders are thrusters.
3
Acoustic Scan
.
1
i
Random errors LMS line fit Specular reflections Double reflections Large angle of incedance Reflections in a Dihedral
Laser range finders
Distance measurement by projection of a sin- gle laser beam. Two types
Triangulation (Simple design but calibration -
problems on a mobile platform)
Time-of-flight (Self-contained unit + no cal- ibration problems, better accuracy but state of the art technology)
Characteristics:
Indoor or outdoor
High resolution images (either 1-D or 2-D)
Fast
High resolution depth measurement
High cost, fragile, powerresolution/speed tradeoff
Sensitive to material properties (e*g* spec- ular materials)
1
Laser range finders: I Triangulation I
.on
Laser I
Resolution/power/time tradeoff
Example: HILARE (LAAS)
s ize : 1.10 x 1.10 x 0.70m w I q h t I LO0 kq
speed : 1 m / s e c max
wheel
caster
i ! *. .. .,...._,.I
! ... *
. . ..
. ..... ..:.
Figure 3 a : Actual data
0, . . . . . . , . . . . . .
i
Figure 3 b : on-line polygonal representation
Figure 1 Figure 3 : Environment circular scanning
with the Laser range-finder
Example: Environment a1 Research Institute of Michigan (ERIM)
Time-of- flight
64 x 256 range images, 30° x 80' Reflectance image
8-bit range fiom 0 to 64 feet (3 inches res- olution)
Similar devices: Odetics, ASV sensor for Ohio State Univ. walking machine.
Range image
Reflectance image
7
Representation of perceptual features for mobile robots
1. Representations of uncertainty and sensor 8- hmon
2.
Special
Terrain
representations :
maps
3. probability maps
1 .
SENSOR FUSION: How to combine observations from different sensors (e.g. passive stereo and active sonar)
Feature observed by three different sensors
Veh
Uncertainty ellipses
MOTION FUSION: How to merge the two observations?
Feature observed from
Feature observed from position 2.
\ Position 2
Uncertainty ellipses
TRADITIONAL APPROACH
Batch of observations (sensor measurements)
\ Resulting uncertainty ellipse
Best estimate of observed feature
K A L W FILTERING
New observation:
Sensor
Current estimate: X c t
New estimate:
c !
Linear Filtering
Where:
z is a I x 1 measurement vector,
X _ is the n x 1 vector to be estimated,
and 2 is a random additive measurement er- ror with I x I covariance matrix R.
The estimate of x is given by:
The best estimate is: T -1 -1 T -1 x- ^ - ( H I ? H ) H R y
Justification: Maximum likelihood approach:
1
Bayesian approach
Best estimation according to minimum vari- ance criterion:
minl(2 - z) T (2 - z)p(zlz)dx
Solution is:
In case of a linear model, the best estimate is:
-1 T -1 -1 T -1 ^ - ( P o + H R H ) H R y - x- Where p0 is the a priori covariance matrix of
2
Recursive filtering
New problem:
We want to find the best estimate zk+l based -on a new measurement zk and the previous es-
Recursive solution:
Where K k is a gain matrix given by: -1 Kk = pkHkT+l(Rk+l +*k+l P k HT k + l )
And the covariance matrix of &+I is updated by:
3
Extended Kalman filter New problem: we have an observation z that
depends on the a parameter vector a by the relation:
f (z, a) = 0 and z = z/ + C, where g is additive noise. Given an estimate a* and a new measurement z what is the new estimate of a?
Linearization of f
Linear measurement equation (as before):
where: df * y = +(&,a*) + -a da- -
and df
4
4 Building maps from matching line segments: ( Faugeras, Ayache, INRIA)
Vehicle position i
I Observations I
Kalman filter Observed set of 3-D line segments , Si, observed at position i.
Updated map of 3-D line segments
-!ication to line matching
xented by the intersection of
x = a z + p
7 0 segments Si ( i = 1,2) supported -3s Li Of parameters (ai, bi,Pii qi ) with 2 matrix (from trinocular stereo) Ai sformation T from frame 2 to frame -ariance matrix A , decide whether:
S? are two instances of the same segment, and
3mpute the uncertainty (Le. the co- 3 J matrix) of the fused segment.
5
The EKF is used to compute:
1 . The estimate of the transformed by T of s2, LQ = (9, b/2,P'2,4 I 2 ), and
2. The covariance matrix computed from
The two lines are matched if
is greater than a threshold s corresponding to a probability of 95% ( x 2 ) distribution.
6
Figure 1 4 Edge segm~mr mpohcd in ~ P I C ~ pair of Figure 12
& t
1
[ )&? j j & J Figure 18: Covpivlce manices a0;rChai to che endpoinu of the rsonzcnrctcd
segments of the offiic mom observed in position 1
Figure 13: Polygonal rppraxkdoa of the edge points of a slcrro pair of h e same offke rarrnobrcmd ia position 2
I
Figure IS: Edge segments maahed in s t ~ r r ~ pair of Figure 13
F i k 17: Haito&rl and nniul p j d c m of the reconstructed segments oftbe dAa mom obrsvsd in position 2
Figure 23: Fusion m position 2 of rhc segments ~consaucted in posifions 1 md 2
ELEVATION M A P S
Sensor
Elevation Uncertainty Slope Curvature
Discrete grid cell (value = traversability cost for path planner)
/ Reference ground plane
Sensor
Sensor
6 Visible regions
Application: cros s-countrv navigation.
(K. O h , D. Payton, Hughes AI Center)
Goal: Real-time perception for cross-country autonomous navigation. Local navigation uses an initial map-based plan.
Vehicle: Martin Marietta 8-wheels vehicle.
Sensor: ERIM laser range finder (delivers 64x256 range images, range is from 0 to 64 feet).
Environment: Outdoor rugged cross-country terrain, including bushes, gullies, rocks, and steep slopes.
.
Perception cycle for cross-country navigation
Obstacle map
Range image from I laser range finder
4
I 1 Sparse elevation map I f I Smoothed dense map 1
I Fused maps
Vehicle model f Trajectories I
Position sensors f
Strategies -4 Navigation 1
Vehicle model:
Normal Bad suspension Slope Clearance
Navigation:
Elevation map
Vehicle model
Predefined trajectories
---.I__..--- I .. . . . . . . . . . .
Figure 4. 3D view of CEM.
Figure 8. CEM from Figure 3 with curved trajec- tories.
Occupancy maps:
( Moravec, Elfes (CMU), Stewart ( M I / Woos Hole))
General description:
The world is represented by a set of regularily spaced cells (grid).
Each cell contains a probability P.
P is the probability that the cell is part of an object.
Each sensor rneaurernent is modelled as a distribution of probabilities over the grid.
The probability at a grid cell is computed by combining the contributions from many sensor measurements.
Is this grid cell occupied? visible? empty?
With what probabhk?
Field of View
Discrete Grid
- Vehicle Locations
3-D OCCUPANCY GRIDS 6 degrees of freedom: R, Tx, Ty, Tz
Sensor
\
Measurement cone
Applications:
Moravec/El.fes/Matthies :
Indoor robot with sonar and stereo cameras. Occupancy map building from both sensors. Used in navigation.
Stewart:
Underwater vehicle with sonar. Full 3-D occupancy map built from many images registered by using the vehicle's position sensors. Extraction of surfaces from the 3-D mid.
U
Application: detailed bottom surface mapping.
Advantages:
Sensor model taken into account.
Feature
Natural
extraction is not necessary.
way of fusing multiple sensors.
.
Computing the probabilities
Problem: Define the probability that a given cell is in state si, given evidence (Le. measure- ment) e. Bayesian model:
P(0CC
P(R~OCC) and P(R(EMP) are given by the sen- sor model, P(OCC) and P(EMP) are the cur- rent probability values in the map. Initially: P(0CC) = P(EMP) = 0.5 (no information).
1
Computing the probabilities
The simplifying assumption: P(R~EMP) = 1 - P ( R I OCC) provides intuitive properties:
Updating independent of the ordering of mea- surements.
~ ( O C C ) = 0.5 (unknown) is a no-op. Conflicting measurements cancel.
2
I
Probabilistic sonar map (from Elfes && Matthies)
Probability
Y'
Measured range b
1 Occupied
I \
/
D
X Sensor
. .
I
Figure3 Probabilistic S o w Sensor Intcrprcution Modcl. Thc probability prcifiie shown cormponds to a d i n g urkcn by a scnsor positioned at the upper left. pointing to thc low- righr The plane shows thc uM(" kvel. Values above the plane rrprescnt 0CtL'pfE.D probabiliucs, and values below npretent E.W probabililiu
Uncertamtv from stereo J
Stereo cameras
P(0CCIR) = P1 + P2 - P1 . a
PI = 0.5 = unknown probability level P2 = maximum allowed P(.OCC) a = 1 .O at edge points, < '1.0 elsewhere
*
Sonar
+
Stereo
iC)
Integra tion
p:i& ......... .. , 22-j . . . . . . . . .... .......... Q 1: : c : : : :
..*.. :::: :*.
. . . . . .
y .. .. ..
Figure 6: Occupancy Maps Csncratcd by Sonar, Stcrco and Scnsor InlCgnlion. OCCUPIED rcgions arc rnarkcd by shadcd squues. EMPTY by do& fading 10 whitc space, and Uh'KNOWh' by + signs.
732
1.
262
Fig. 13. shown
IEEE JOURNAL OF ROBOTICS AND AUTOMATION. VOL. RA-3, NC. 3. JUNE 1987 ’
.
.......... .......... .......... ........,.
Example run. This run was performed indoors, in Mobile Robot Lab. Distances arc in ft. Grid size is 0.5 ft. Planned path is as dotted line. and route actually followed by robot as solid line segments. Starting point is solid + and goal, solid x .
Detailed description of an autonomous vehicle: the NAVLAB
Self-contained vehicle for:
Road following with or without map.
- Object detection.
Map building.
using:
One color camera.
One laser range finder.
On-board computing (Suns).
Controller.
1
I z w
I
Camera
polyhedron r v l
Object Recognition
.
Motion Planning
-
landmark
tree
Tokens Tokens
I Blackboard Manager
I I
0
0
0
Moving Coordinates
Time
Distributed Processing
Environment
Paved curved roads.
Non-uniform road appearance.
Changing conditions (illumination, weather. . .etc).
Discrete obstacles.
Strong shadows from obstacles.
2
Components of the road following algorithm
Color classification: The color of the road is significantly different than the background.
Texture computation: The sides of the road are usually more textured.
.Road location in image: The geometry of the road must be determined.
Calibration: The road detected in image space must be converted to the vehicle’s co- ordinate system in order to steer the vehicle.
3
Image
Non-road Model Appearances
(RGBT)
f L
Roamon-road 4 Classification 1
v Road b osition Hou hfor
& Orientation
Road Model Width
Position Orientation
Surface Appearance (RGBT)
I
Self Clustering &
Update I
- ' I I Vehicle Motion
.
Color classification
The color is divided into n classes. Each pixel is classified into one of the classes using the distance:
Where:
x is the (red,green,blue) vector at the current pixel.
mi is the mean (red,green,blue) value of class 1.
ci is the covariance matrix of class i.
The set of classes is divided into road classes and non-road classes.
4
Intensity image (480x5 12)
high-frequenc y gradient (240x256)
Low-frequenc y gradient (60x64)
1 Average I
I -
value (120x128) 1
Normalized gradient hf gradient
I a If gradient +p .average value (480x5 12)
Binary image of micro edges (480x5 12) -
I
1 Image of edge density (3 0x3 2)
I
I I Texture classification Road/Non-Road
Combining texture and color
For each class and each pixel:
Pi = (1 - a)Pi T +aPi C
Where -
Pi = confidence that the pixel belongs to class i
m
~ f ' = confidence that the pixel belongs to class i based on texture
class i based on color
-.
.Pc I = confidence that the pixel belongs to
Final confidence for each pixel:
C = ma~(Pi, i f Road) - m=(Pi, i E NonRoad)
c > 0 e the pixel is classified as road.
5
Road representation
Vanishing line
shing point
Hough transform
I
Horizon point
Voting algorithm
Quantized 2-D parameter space (P , e ) (32 lev-
An image point (row, C O ~ ) can be votes for the els for P, 20 levels for @)
set of road location:
(P , e ) = (COZ + (row - rhorizon ) x tan8,o)
Each pixel casts votes proportional to the road/nQn-road classification confidence.
The maximum in ( p , e ) space is the reported road location.
6
.. Red operator Reduce - - -
Green b Edge -
- - Red -
n 4 Yrocessing cvcle U d
Color classification
Texture image H Color I
Texture classifka tion
statistics u 1 Grass Classification Fl combination Sample
colors 1 Grass 1 1-0 Grass 0
Road 1 I,, edges I I
i-' Road 0 1
r Highest Hough
space vote -
Simple calibration procedure Measure the distance between rows ri and the vehicle.
Detected road
Distance between Qi and the center of the vehcle =
(Ci - ci)/ppmi
ppmi = pixels per meter at row ri.
Dynamic road following
Statistics from
Updated statistics and road location
Vehicle steering
Predicted vehicle position (t seconds later)
0
Some lessons from road following
More accurate calibration is needed.
More dynamic range in the color cameras is required to handled a wide range of con-
* * . ditions. More detailed road model (intersections, curved roads..).
More flexible color classification.
7
Range data
Sensor:
Time-of-flight laser range finder (ERIM) 64 x 256 range images, 30° x 80°
Reflectance image
%bit range from 0 to 64 feet (3 inches res- olution)
Purpose: -
Obstacle detection f- path planning around discrete objects
Terrain modeling +- path planning across open terrain
3-D map building c- exploration, incremen- tal description improvement
1
Sensor Obstacle detection
I/ X Ranrreimaee
Discrete elevation map
Polygonal map of obstacles
3-D map building
Shadow
I
Image 1
Motion T
Image 2
I I +&- Prediction window from T and image 1
Range processing cycle
image
Elevation map
Current map
Obstacle map
Vehicle
Terrain regions
Terrain modeling
The part of the field of view that is not part of any object can be segmented into regions: 1. Find edges in the elevation map
2. Find seed points in the elevation map
3. Apply region growing based on smoothness constraint
4. Report the region (e.g. as polygons . - . .
2
Putting everything together
IT< Perception modules L K
Driving u&s/” Steering commands
Re-loaded map
Additions to the current system
Uncertainty representation for road location, vehicle position, and objects’ locations. Up- dating through filtering
Better path planner, including off-road ca- pabilities
0 Improved vision: two cameras, intersections..
Map revision . . . . . . . . _. . . . . . . .
1
Years left Device Difficulty
0 Toys cost 1 12 Watchdog re liabi I ity (Brooks, Durrant-White) ?? Factory transport planned motion 1 (Crowley) Industrial cleaning coverage 1 O(Chatil1a)
73 a . Wheelchair plan ning, I iabi lity 1993 (Harmon) Tank simulator rough terrain 77 e . Mine sweeper coverage 5 (Somalvico) 30 (Binford) 10 (Chang) Street sweeper traffic
Household servant tough manipulation
Mail delivery manipulation Garbage collection manipulation
77 . . Tank weapons 10 (Harmon) Construction force (Somalvico) 7 (Graefe) Chauffeur psychology
2
7
A partial list of research efforts in mobile robotics (from J.L. Crowley, "The State of
the Art in Mobile Robotics") LAAS-C~W, T. Ave. Colonel Roche, 31077 Toulouse, France. Project Hilare, Principle investigators: George Giralt, Raja Chatila, J-P. Laumond. Perhaps one of the longest running projeczs. Reccnt applications include a cleaning robot for the Paris Metro as well as a ware-house robot
SRI-Internationat, AI Group, 333 Ravenswood Ave., Menlo Park, California, 94025 USA. Principle Investigators: Stan Rosenschein and Leslie Kaibling. Robot hardware: Flakey the robot, designed and built by Stan Reifle.
C-MU Robotics Institute, Schenley Park, Pittsburgh Pa, I52 13 USA. Principle investigators: Takeo Kana&, Chuck Thorpe, Hans Moravec, Red Whittaker. At least 5 distinct projects have been performed at C-MU in the last 5 years. Major projects include The ALV NavLab, the Terragator, the Denning surveillance robot, the W, and Moravec's Pluto, Neptune and Uranus.
MIT AI Lab: 545 Technology Square, Cambridge MA, 02139, Principle Investigator. Rod Brooks. Robots: Allen, Torn, Jerry, Sydney, Seymour.
. .
INRIA: Domaine de Voluceau, Rocquenco- BP 105, 78i53 Le Chesnay, France. Principal Investigators: Olivier Faugeras, Nicholas Ayache, Francis Lustman. A commercial copy of the INRIA robot is sold by RobotSoft SARL, of Asniers France.
LIFIA: W G , 46 ave Felix Viallet, 38031 Grcnoble, France. Principal Investigator J i m Crowley. Developing a Surveillance Robot for project EUREKA - Mithra. Currently using a Denning robot named Lurch.
GM Research, Warren Michigan, USA: Principle Investigators: Steve Holland, Bob Tilove
Univ. of Amsterdam, Kmslaan 4090, The Netherlands. Principle Investigator Dr. Willem Duinker.
Stanford University: AI Laboratory, Computer Science Dept, Stanford University, Stanford, Ca, 94305 USA. Principal Investigator: Thomas Binford.
. .
ORNL: P.O. Box X, Oak Ridge Tenn, 37831 USA. Principle Investigator: Charles Weis bin.
FMC Corp. 1205 Coleman Ave., Santa Clara, Ca 95052. Principle Investigator Andrew Chang Military applications of mobile robots.
NBS: Inausmal Systems Division, National Bureau of Standards, Building 220/B 124, Gaithersberg, MD. 20899. prinicipal Investigator: Marty Herman.
Insitiit der Bundeswehr MQnchen, Inst fur Nfesstechnik, 8014 Neubiberg, W. Germany. Principle Investigator: Volker Graefe. Real time road-following system inte,mting simple real time vision with control theory.
Jet Propulsion Laboratory, 4800 Oak Grove Drive, Pasadena, Ca. 9 1 109 USA. Principle Investigator Brian Wilcox.
2
References
R.C. Arkin. E. Riseman, A. Hanson. Visual strategies for mobile robot navigation. In Proc. Workshop on Computer Vision. 1987.
N. Ayache, F. Lusunan. Fast and reliable trinocular stexeovision. In Proc. ICCV'87. 1987.
A. Bergman, C. K. Cowan. Noise-Tolerant Range Analysis for Autonomous Navigation. In IEEE Conf. on Robotics and Automarion. San Francisco, 1986.
R. Brooks. Aspects of mobile robot visual map making. In Second International Robotics Research Symposium. MlT press, 1985.
R. Brooks, AM. Flynn, T. Marill. Self calibration of motion and stefeo vision for mobile robot navigation. In Proc. Image Understanding Workshop. Cambridge, 1988.
M.K. Brown. Locating object surfaces with an ultrasonic range sensor. In Proc. IEEE International Conference on Robotics and Automation. St Louis, 1985.
S.S. Chen. Multisensor fusion and navigation of mobile robots. International Jownal Intelligent Systems, Vol. 2 , 1987.
J. Crowley. Dynamic world modeling for an intelligent mobile robot using a rotating ultrasonic ranging device. In Proc. IEEE International Conference OR Robotics and A1Uomotion. St Louis, 1985.
M. J. Daily, J. G. Harris, K. Reiser. Detecting Obstacles in Range Imagery. In Image Understanding Workshop. Los Angeles, 1987.
MJ. Daily, J.G.Harris,K. Reiser. An operational perception system for cross-country navigation. In Proc. Image Understanding Workshop. Cambridge, 1988.
E.D. Dickmanns, A. Zapp. A curvature-based scheme for improving road vehicle guidance by computer vision. In Proc. SPIE Conference 727 on Mobile Robots. Cambridge, 1986.
E.D. Dickmanns, A. Zapp. Autonomous high-speed road vehicle guidance by computer vision. In Proc. 10th IFAC. Munich, 1987.
K.C. Drake, E.S. McVey, R.M. Tingo. Experimental position and rangmg results for mobile robots. Jownal of Robotics and Automation, Vol. 3 ,1987.
M. Drumheller. Mobile robot lacalization using sonar.
R. T. Dunlay, D. G. Morgenthaler. Obstacle detection and avoidance from range data. In Proc. SPIE Mobile Robots Conference. Cambridge, MA, 1986.
PtMI-9,1987.
3
~ 7 1
A. Elfes. Sonar-based real-world mapping and navigation. Journal of Robotics and Automation, Vol. 3 , 1987.
O.D. Faugeras, F. Lustman, G. Toscani. Motion and structure from motion from point and line matches. In Proc. I C W 8 7 . 1987.
O.D. Faugeras, N. Ayache, B. Faverjon. Building visual maps by combining noisy stereo measurements. In Proc. IEEE Conf. on Robotics and Automation. 1986.
G. Gimlf R. Chatila, M. Vaisset. An integrated navigation and motion control system for autonomous multisensory mobile robots. In Proc. 1st International Symposium Robotics Research. Cambridge, 1984.
K.D. Gremban, C.E. Thorpe, T. Kanade. Geomemc camera calibration using systems of linear equations. In Proc. Image Understanding Workshop. Cambridge, 1988.
S.Y. Harmon. The ground surveillance robot (GSR): An autonomous vehicle designed to transit unknown terrain. Journal of Robotics and Automation, Vol. 3 , 1987.
M. Hebert, T. Kanade. First results on outdoor scene analysis. In Proc. IEEE Robotics and Aufomotion. San Francisco, 1985.
M. Hebert T. Kanade. 3-D vision for outdoor navigation by an autonomous vehicle. In Proc. Image Understanding Workshop. Cambridge, 1988.
T. Kanade, C. Thorpe, W. Whittaker. Autonomous Land Vehicle at CMU. In Proc. ACM Computer Conference. Cincinnati, February, 1986.
D J. Knegman, T.O. Binford. Generic models for robot navigation. In Proc. Image Understanding Workshop. Cambridge, 1988.
D. Knegman, E. Triendl, T.O. Binford. A mobile robot: sensing, planning and locomotion. In Proc. IEEE Con5 on Robotics and Automation. 1987.
D. Kuan, G. Phipps, A.C. Hsueh. Autonomous land vehicle road following. In Proc. ICCV. 1987.
R. Kuc, M.W. Siegel. Physically based simulation model for acoustic sensor robot navigation.
S.P. Liou, R.C. Jain. Road following using vanishing points. Compufer Vision, Graphics, and Image Processing , 1987.
L. Matrhies, S. A. Shafer. Error Modelling in Stereo Navigation. Technical Report CMU-RI-TR-86- 140, Carnegie-Mellon University, the Robotics Institute, 1986.
L. Matthies, S.A. Shafer. Error modeling in stereo navigation. Journal of Robotics and Automation, Vol. 3 , 1987.
PAMI-9 , 1987.
4
1341
[381
1401
1411
1431
[441
L. Matthies, R. Szeliski, T. Kanade. Kalman filter-based algorithms for estimating depth from image sequences. In Proc. Image Understanding Workshop. Cambridge, 1988.
D. Miller. A spaual representation system for mobile robots. In Proc. IEEE International Conference on Robotics and Automation. St Louis, 1985.
H.P. Moravec, A. Elfes. High-resolution maps from wide angle sonar. In Proc. IEEE International Cogerence on Robotics and Automation. St Louis, 1985.
R.C. Nelson, J. Aloirnonos. Using flow field divergence for obstacle avoidance in visual navigation. In Proc. Image Understanding Workshop. Cambridge, 1988.
K. E. Oh, F. M. Vilnroaer, M. J. Daily, K. Reiser. Developments in Knowledge-Based Vision for Obsracie Detection and Avoidance. In Image Understanding Workshop. Los Angeles, 1987.
J.L. Olivier, F. Ozguner. A navigation algorithm for an intelligent vehicle with a laser rangefinder. In Proc. IEEE Conf. on Robotics and Automation. San Francisco, 1986.
S.V. Rao, S .S. Iyengar, C.C. Jorgenson, CX. Weisbin. Concurrent algorithms for autonomous robot navigation in an unexplored terrain. In Proc. IEEE Cog. on Robotics and Automation. San Francisco, 1986.
AX. de Saint Vincent. A 3D perception system for the mobile robot HILARE. In Proc. IEEE Conf. on Robotics and Automation. San Francisco, 1986.
K. Sakwai, H. Zen, H. Ohm Y. Ushioda, S. Ozawa. Analysis of a road image as seen from a vehicle. In Proc. ICCV. 1987.
S. A. Shafer, A. Stem, C. E. Thorpe. An Architecture for Sensor Fusion in a Mobile Robot. Technical Report CMU-RI-TR-86-9, Carnegie-Mellon University, the Robotics Institute, 1986.
R.C. Smith, P. Cheeseman. On the representation and estimation of spatial uncertainty. International Journal of Robotics Research , 1986.
W.K. Stewart. A non-deterministic approach to 3-D modeling underwater. In Proc. Fifth International Symposium on Unmanned Untethered Submersible. University of New
Hampshire, 1987.
C.E. Thorpe. The CMU rover and the FIDO vision and navigation system. PhD thesis, Camegie-mellon University, 1984.
C.E. Thorpe, M. Hebert, T. Kanade, S.A. Shafer. Vision and navigation for the Camegie-Mellon Navlab. PAMI lO(3). 1988.
D. Y. Tseng, M. J. Daily, K. E. Oh, K. Reiser, F. M. Vilnrotter. Knowledge-based vision techniques annual technical report. Technical Report ETL-0431, U.S. Army ETL, Fort Belvoir, VA. 1986.
S. Tsuji. J.Y. Zheng. Visual path planning for a mobile robot. In Proc. IJCAI. 1987.
4 5
[48] R. Wallace, K. Matsuzaki, Y. Goto, Y. Webb, J. Crisman, T. Kanade. Progress on Robot Road Following. In IEEE Con$ on Robotics and Automation. San Francisco, 1986. A.M. Waxman, JJ. LeMoigne, L.S. Davis, B. Srinivasan, T.R. Kushner, E. Liang. T. Siddalingaiah. A visual navigation system for auatonomous land vehicles. Journal of Robotics and Automation, Vol. 3 , 1987.
W.J. Wolfe, N. Marquina, Eds. Mobile Robots II. In Proc. SPIE 852. Cambridge, MA, 1987.
[49]
[SO]