Performance evaluation of particle filter based visual tracking

摘要

Tracking is a fundamental problem for event recognition. Particle filter (PF) is acknowledged for its efficiency in dealing with multi-modal visual tracking problem of general nonlinear and non-Gaussian system. It therefore emerged as an appealing tool for tracking objects in video sequences. Although many proposals have been put forward to deal with various scenarios and enhance PF convergence properties, it is acknowledged that comprehensive evaluations of the proposals are still lacking. This paper aims to contribute to this ongoing research. Especially, simulated videos were created to analyze the influence of the target appearance according to various noise intensities that entails partial or full occlusion scenarios. In the simulated videos, the experiment provided more accurate conclusions given the range of involved factors, w.r.t target model, similarity measurement, and environmental distraction. To validate the conclusion from the simulated videos, the experiment was also conducted in the benchmark videos.