Performance evaluation of multitarget tracking algorithms

Date of Completion

January 2002

Keywords

Engineering, Electronics and Electrical

Degree

Ph.D.

Abstract

When tracking multiple targets using multiple sensors, the performance evaluation of different estimation and data association algorithms becomes critical in system design and global level coordination. One of the most important problems in tracking multiple targets is the measurement origin uncertainty, i.e., one has to deal with missed detections and false alarms when trying to assign uncertain origin measurements to tracks. ^ To better understand the theoretical bound of the parameter estimation of a single target in the presence of clutter, the maximum likelihood estimator with probabilistic data association (ML-PDA) is formulated and the Cramer Rao lower bound (CRLB) is obtained. It is shown that a scalar information reduction factor (IRF) quantifies the information loss due to measurement origin uncertainty. To generalize the ML-PDA technique to the parameter estimation of multiple targets, one needs to decide the number of targets in the observation space based on a certain model selection criterion. A new approach, using the minimum description length (MDL) criterion, is presented for the detection and initiation of multiple tracks with the ML-PDA estimator. ^ The tracking performance using multiple sensors is studied on two data processing algorithms, namely, the centralized filter and a distributed scheme relying on track-to-track fusion. Based on a simple target kinematic model, it is shown that the optimal track-to-track fusion is indeed suboptimal compared with the centralized solution and the performance difference between the centralized and distributed trackers increases as the number of sensors increases even under the assumption of perfect data association. ^ Finally, data association with covariance modification is studied under the top-M multiple hypothesis tracking (TMHT) framework. The N dimensional assignment problem is formulated including possibly unresolved measurements due to the finite resolution of a sensor when tracking closely-spaced targets. The M-best solutions of the N dimensional assignment are used to modify the covariance of each track based on the probabilistic data association filter (PDAF) at the rear end of the sliding window. A suboptimal approach using a linear program to solve the relaxed assignment problem is also compared with single best assignment solution and TMHT. ^

Share

COinS