Adaptive pedestrian tracking via patch-based features and spatial-temporal similarity measurement

The existing pedestrian tracking algorithms are still challenging under abrupt illumination, occlusions, out of field of view, and cluttered backgrounds. In this paper, we propose a novel tracking framework to handle the above tracking failures. The framework consists of two steps: image shadow removal and tracking by association. Shadow removal tackles tracking failures under large illumination variation between shaded and non-shaded area. Tracking by association is based on particle filter. Inspired by the observation that the different parts in one pedestrian target have different stable properties, we represent the target with perceptual hashing (PH) and color self-similarity at patch level, and design an adaptive patch-based similarity measurement which simultaneously considers the spatial and temporal appearance information of the target. Quantitative and qualitative evaluations on several sequences show that our method consistently outperforms tracking methods compared.