Reviews
Stanford / VW Autonomous Driving
Last week I was half invited and I half invited myself to attend a meeting for the Stanford Autonomous Vehicle Racing Team along with my VWoA-ERL coworkers. Some interesting ideas were mentioned, such as a method for re-lighting a scene using a light scattering formula to reconstruct points’ surface normals and textural qualities from velodyne data captured by Junior (by Jesse Levinson). At the end of the meeting I offered some of my time to finish a system they’ve been developing for detecting traffic lights for their next-generation Junior vehicle! That should make April and May pretty interesting!
Tracking and Atypical Motion Detection
A review of Dr. Pascal Fua’s talk at UC Berkeley (Jan-5-2009).
I stopped by Cal yesterday to catch a talk by Dr. Fua (Computer Vision Lab, EPFL, Switzerland) on his group’s latest findings resulting from a paper they published last year regarding their Probabilistic Occupancy Map. Using a combination of some very simple and intuitive trajectory models, some dynamic and some linear programming, they have come up with an algorithm which robustly tracks a small number of people in the field of view of several cameras. Alone, this wouldn’t be news, as automated tracking is becomming more common-place in industry (see SportVision, for example). But with their discrete grid on the ground-plane paradyme, they have managed to overcome several key obstacles.
First, some background. The system determines the probability that a person is standing on a specific grid square by accounting for the background subtracted image having a human sized blob just above the square. Then the previous probability that a person’s trajectory would lead them to this grid square is taken into account. Several independent trajectory models are used to come to a more accurate conclusion. Once each probability from each model has been multiplied together, a very clear decision can typically be made as to the location of the person.
Because the trajectory models are given only so much importance compared to the background subtraction, there is little worry about failure to track due to divergence.This also helps a great deal with occlusions. Since blob size changes in direct proportion to distance from each camera, a very specific location can be placed on each subject in each frame and when one or several people walk in front of another person, the system keeps good track of who is who.
The other interesting piece of information this system is good at delivering is labeling those people in its field of view whose motion doesn’t fit any of the trajectory models. This is useful for those concerned with security in places with obvious destinations such as airports, train/metro stations, etc.
There are some problems they are currently working out: The trajectory models seem to work well only under the right circumstances (they are somewhat environment specific at the moment). The system hasn’t yet been implemented in such a way as to keep proper track of two or more people who are holding hands or are very close together for a time. It will either track the two as a single person or swtich their identities. This isn’t so much a problem as an element to the algorithm that has been deliberately ignored, since its cure is obvious. They are looking into using a method other than background subtraction (maybe the human detection model devised at Cal?) to determine who is who without so much depending upon frame to frame probability.
For a look at their code, check out this site.