Multiple pedestrian tracking under first-person perspective using deep neural network and social force optimization

Multiple pedestrian tracking in the first-person perspective is a challenging problem, obstacles of which are mainly caused by camera moving, frequent occlusions, and collision avoidance. To solve the mentioned issues, we proposed a novel deep learning-based approach. Firstly, a dense connection and attention based YOLO (DCA-YOLO) is proposed for ameliorating the detection performance. Then, the detection results are sent to a wide residual network for feature extraction. We use the Kuhn-Munkres algorithm to construct a similarity matrix and find the best match of two detection boxes. To tackle the frequent occlusion and ID-switch issues caused by collision avoidances or grouping behavior, we introduce a social force model into the proposed network to optimize the tracking results. The experimental results on widely used challenging MOT2015 and MOT2016 benchmarks demonstrate the effectiveness of our proposed algorithm.