In multiple pedestrian tracking, which algorithm can detect fast and robust?
in crowded senses with occlusion.
Hello. If you are tracking multiple objects from a video sequence you will encounter similar problems to what typically occurs in radar-based multiple target tracking such as the presence of multiple, closely-spaced objects, some of which may change direction (i.e. manoeuvre). You will also have occlusion to deal with, as you know. There is no magic arrow or best algorithm and every situation is different. Algorithms all have their strong and weak points. Your requirement that the algorithm be both fast and robust is somewhat contradictory as these criteria are in conflict.
You can try anything from the simplest nearest-neighbour Kalman filter to multiple hypothesis trackers. If you have additional information on the objects, such as their size or colour, etc, this may help to resolve ambiguities, but the basic problem has been well studied.
A survey of conventional approaches is included in the link. More recent approaches are based on particle filtering and PHD filtering, which do not always take object identity (labels) into account. At the present time it is not possible to say if they are superior to MHT or multiple scan assignment, as there are few if any objective comparisons.
Article Taxonomy of multiple target tracking methods
Hello,
I would like to suggest you this other reference that explain how this issue is addressed for pedestrian tracking:
http://rd.springer.com/article/10.1007/s11263-005-4797-0
The hypothesis used here is much more complex than the linear one indicated by Kirill Kondrashov but it may provide better results.
Hello. The previous poster's reference to "discrete choice models" is interesting. It seems to be a decision process model for pedestrian motion. My main question about this is how you can pose an estimation problem on a what is effectively a hybrid system (multiple discrete modes and continuous dynamics) as a control problem? That is to say, the pedestrian's "utility function" is not observable by the video sensor so how can you figure out what their sequence of decisions is? (Hybrid systems use a set of possible motion models, one of which may apply at a given time, with a Markov chain model for the switching from one to another.)
A more conventional approach would be to use any one of a number of estimation techniques for hybrid systems or manoeuvring target tracking methods. Examples include interacting multiple model (IMM), generalised pseudo-Bayesian (GPB) or likelihood-based multiple model approaches. The link below describes some of these. The second link has some material on mixture models for pedestrian motion.
Technical Report A Survey of Manoeuvring Target Tracking Methods
Conference Paper Inertial Navigation Versus Pedestrian Dead Reckoning: Optimi...
- Badges
- Science topic
- Similar topics
- Biological Science
- Microbiology
- Virology
- Virus