Dynamic feature and signature selection for robust tracking of multiple objects

Abstract

The goal of this paper is to introduce a new tracking framework, which exploits dynamic feature and signature selection techniques for data association models. It performs robust multiple object tracking in a noisy, cluttered environment with closely spaced targets. This method extends the back-end processing capabilities of tracking systems by creating a hierarchy between the parallelly extracted features. These features are dynamically selected based on spatio-temporal consistency weight function, which maximizes the robustness of data association, and reduces the overall complexity of the algorithm.