Lightweight Particle Filter for Robust Visual Tracking

Particle-filter-based methods have made significant contributions to achieve robust visual tracking over the past few years. Although these tracking algorithms achieve superior performance, their real-time application has typically been hindered by the high computational burden attributed to the sampling process of the particle filter. In this paper, we propose an innovative lightweight particle filter tracking (LPFT) approach that retains the robust tracking ability of the particle filter while alleviating the time-consuming sampling burden by the correlation filter method with response maps. Specifically, the proposed tracker fully utilizes the location and confidence information of these maps to predict the object, thereby training a more robust correlation filter based on more accurate predictions compared with standard correlation filters. Additionally, in contrast to most particle-filter-based trackers, the LPFT tracker capable of handling large-scale variations can be incorporated into any tracker that requires fast and accurate scale estimation. We propose the generic parallelization framework and implement another version of the LPFT tracker to validate this incorporation. Extensive evaluations on the OTB-2013 and OTB-2015 benchmarks demonstrate that the proposed tracker achieves very competitive performances with state-of-the-art trackers while operating at over 60 frames per second using hand-crafted features.