A Modular Approach to the Analysis and Evaluation of Particle Filters for Figure Tracking.pdfVIP

A Modular Approach to the Analysis and Evaluation of Particle Filters for Figure Tracking Ping Wang and James M. Rehg Georgia Institute of Technology GVU Center, TSRB 85 5th St. NW, Atlanta GA 30332 {pingwang, rehg}@ Abstract This paper presents the first systematic empirical study of the particle filter (PF) algorithms for human figure track- ing in video. Our analysis and evaluation follows a modu- lar approach which is based upon the underlying statisti- cal principles and computational concerns that govern the performance of PF algorithms. Based on our analysis, we propose a novel PF algorithm for figure tracking with superior performance called the Optimized Unscented PF. We examine the role of edge and template features, intro- duce computationally-equivalent sample sets, and describe a method for the automatic acquisition of reference data using standard motion capture hardware. The software and test data are made publicly-available on our project web- site. 1. Introduction The problem of tracking the human figure in video is commonly treated as a sequential state density estimation problem and solved using a version of the Particle Filter (PF) [3, 4, 5, 24]. The theoretical basis for the PF derives from the statistical theory of sequential monte carlo meth- ods. The statistical literature on PF includes mathemati- cal derivations of PF [7, 13], proof of convergence, error bound [2, 14, 27], optimal choice of functions and param- eters [7, 13, 14], and measurement of effective sample size [13]. While empirical comparisons of the PF algorithms ex- ist in the statistical literature [1], they are typically restricted to low-dimensional problems, which greatly limits their ap- plicability to figure tracking. In figure tracking the state space routinely exceeds 20 dimensions, making it a very challenging domain for the PF. Since the number of samples needed for effective track- ing is exponential in the dimension of the state space, the basic PF algorithms such as