Abstract.
Two basic facts motivate this paper: (1) particle filter based
trackers have become increasingly powerful in recent years, and (2) object
detectors using statistical learning algorithms often work at a near realtime
rate.
We present the use of classifiers as likelihood observation function of a
particle filter. The original resulting method is able to simultaneously
recognize and track an object using only a statistical model learnt from
a generic database.
Our main contribution is the definition of a likelihood function which
is produced directly from the outputs of a classifier. This function is an
estimation of calibrated probabilities P(class|data). Parameters of the
function are estimated to minimize the negative log likelihood of the
training data, which is a cross-entropy error function.
Since a generic statistical model is used, the tracking does not need
any image based model learnt inline. Moreover, the tracking is robust
to appearance variation because the statistical learning is trained with
many poses, illumination conditions and instances of the object.
We have implemented the method for two recent popular classifiers: (1)
Support Vector Machines and (2) Adaboost. An experimental evaluation
shows that the approach can be used for popular applications like
pedestrian or vehicle detection and tracking.
Finally, we demonstrate that an efficient implementation provides a realtime
system on which only a fraction of CPU time is required to track
at frame rate.
1 Introduction
We address the problem of real-time detection and tracking of an object using
only a generic statistical model of the object. The idea is to bring together two
popular fields of computer vision: statistical learning algorithms and particle
filtering. Statistically based object detector using boosting [1] and support vector
machine (SVM) [2] are now fast enough to run in real-time. Furthermore, particle
filter based trackers [3, 4] provide successful solutions in following objects in
clutter from a video. They have been used with edge-based [4], appearance [5] or
kinematic [6] models, most of them, learnt for the specific object to be tracked.
We propose to use a generic model of the class of the object, computed offline
by a statistical learning algorithm from a database.

Download full report
http://googleurl?sa=t&source=web&cd=1&ve...hierry.pdf&ei=mqRUTvD_CYOyrAf2j-mvDg&usg=AFQjCNGEl34UwivaKcx0VCDPsMj8vysuqQ&sig2=fcztUrdlj3la7Y87jG_qOA