Adaptive Fusion of Sub-band Particle Filters for Robust Tracking of Multiple Objects in Video

Abstract

Video tracking is a relevant research topic because of its many surveillance, robotics, and biomedical applications. Although remarkable progress was made on this topic the capability to track objects precisely in video frames that contain difficult conditions, such as an abrupt variation in scene illumination, incomplete object camouflage, background motion and shadow, presence of objects with distinct sizes and contrasts, and presence of noise in the video frame, is still considered a vital research problem. To overcome the presence of these difficult conditions, we proposed a robust multi-scale tracker that used different sub-bands frame in the wavelet domain to express a captured video frame. Then N independent particle filters are employed to a selected subset of these sub-bands, where the selection of this wavelet sub-bands varies with every captured frame. Finally, the output position paths of these N independent particle filters were fused to obtain more precise position paths for moving objects in the video. To show the robustness of the proposed multi-scale video tracker, we employed it to various example videos that have different challenges. Opposed to a standard full-resolution particle filter-based tracker and a single wavelet sub-band (LL)2 based tracker, the proposed multi-scale tracker shows greater tracking performance.