Multitarget tracking through conditional probability estimation

Chad M Aeschliman, Purdue University

Abstract

Multitarget tracking is a challenging problem with a wide range of applications. Often this problem is approached through a tracking-by-detection framework in which a classifier is trained to distinguish the target of interest from all other objects. The classifier is often trained offline, which limits the domain of possible targets and generally requires a large and diverse set of training images. In this thesis, we focus on online training of a probabilistic classifier which estimates the class conditional probability for the target of interest. Online estimation of conditional probability for multitarget tracking poses several challenging issues due to the high-dimensionality of the input and the small number of training samples. We explore two fundamentally different solutions to this problem. The first is based on building a generative model of the target class probability followed by applying Bayes' rule to obtain the conditional probability. This approach is limited to situations in which the background is stationary. However, this approach builds a full model of the target allowing individual targets to be segmented out efficiently for improved robustness to occlusion and overlapping targets. The second approach is based on directly training an estimator for the conditional probability using an efficient linear programming strategy. This approach is more broadly applicable since it places no limitations on the background. However, it is limited to fixed bounding-box based tracking which can lead to drift. Extensive experimental results indicate that both approaches yield state-of-the-art performance in various multitarget tracking scenarios.

Degree

Ph.D.

Advisors

Kak, Purdue University.

Subject Area

Computer Engineering

COinS