Multiple target tracking with application to image sequence processing

Wai Ying Kan, Purdue University


This research addresses an application of video image processing and a simple target tracking algorithm to a highway surveillance problem. Two different vehicle detection algorithms, the first based on a Gaussian mixture model and the second based on a hidden Markov model, are described. The image pixels are modeled by Gaussian mixture in the first formulation. The primary goal of the first detection work is the development of detection thresholds computed from estimated lane-image statistics which will automatically adapt to changing ambient illumination. In the second formulation, vehicle signatures are described using a one-dimensional hidden Markov model for the cross-lane segmented images. Vehicle detection and classification is accomplished by the MAP estimation of the trajectory through the chain. Emphasis is put on the low complexity issue in both algorithms. Simple tracking algorithms based upon nearest neighbor filtering do not correctly consider measurement origin uncertainty and, therefore, fail to perform well in situations of high target density and clutter. To address these issues, a more sophisticated recursive tracking algorithm was developed for approximating the optimal Bayesian tracking estimator in the MSE sense. Assuming independent Gauss-Markov models for the individual targets, then, conditioned on the number of targets, the posterior density of the states given past observations is a Gaussian sum. The number of terms in the sum is determined by the number of ways to associate observations and targets. The complexity of this combinatorial problem results in an optimal filter of exponentially growing memory. Approximation is done by naturally partitioning and grouping those target state estimates into approximate sufficient statistics. A new criterion function is introduced in this approximation process. Both the well-known Probabilistic Data Association filter (PDAF), and its multiple target version, the Joint PDAF, turn out to be special cases of the new algorithm. Comparisons are made between the proposed estimator and the PDAF as well as the Joint PDAF. Improvement is observed for the new estimator.




Krogmeier, Purdue University.

Subject Area

Electrical engineering|Civil engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server