Pedestrian Tracker Utilizing Sensor Fusion and a Gaussian-Mixture Probability Hypothesis Density Filter

This paper presents and evaluates a real-time implementation of Gaussian-mixture probability hypothesis density (GM-PHD) filter aimed at tracking pedestrians in vehicular applications. The GM-PHD is a subsequent development of the probability hypothesis density (PHD) filter first introduced by Mahler as a solution to the multi-target tracking dilemma including computational complexity, data association, and inconsistencies in sensors. An adjusted Gaussian model was utilized to track the measurements persistently in real-time. By fusing sensor information from a frequency modulated continuous wave (FMCW) radar and 2D scanning lidar our system focused on implementing real data for physical implementation in autonomous vehicles. Experiment results document the performance of the GM-PHD framework and exemplify its excellent performance in tracking dynamically changing multi-targets with obfuscations in a real-time scenario.