Online and Real-Time Tracking with the GM-PHD Filter using Group Management and Relative Motion Analysis

摘要

In this paper, we propose an online and real-time multi-target tracking method exploiting the tracking-by-detection approach. The proposed method includes a two-stage data association strategy with the Gaussian mixture probability density filter and an occlusion handling method using group management and motion analysis. Also, we devise a new measure namely sum-of-intersection-over-area to determine the targets' merge, occlusion, and split used in the group management scheme. To verify that proposed framework works efficiently at multi-target tracking tasks, we evaluate our tracker on the UA-DETRAC dataset which contains about 140,000 of images with the vehicle detection responses. The experiment results show that our tracker not only runs faster than 400 fps but also achieves the competitive tracking performance with the second PR-MOTA score against the state-of-the-art trackers.