Dynamic Dual-Peak Network: A real-time human detection network in crowded scenes

Abstract

Human detection in crowded scenes is challenging since the objects occlude and overlap each other. Compared to general pedestrian detection, there is also more variation in human posture. This paper proposes a real-time human detection network, Dynamic Dual-Peak Network (DDPNet), which specifically addresses human object detection in overlapping and crowded scenes. We design a deep cascade fusion module to enhance the feature extraction capability of the anchor-free model for small objects in crowded scenes. In the meantime, the head–body dual-peak activation module is used to improve the prediction score of the central region of the occluded individual through low occlusion components. By this improvement strategy, the network’s ability is enhanced to discriminate individuals in crowded scenes and alleviate the problem caused by individual posture variation. Ultimately, we propose a novel Exhale–Inhale method to adjust the feature mapping ranges for various scale objects dynamically. In the process of ground truth mapping, the overlapping of individual feature information is reduced. Our DDPNet achieves competitive performance on the CrowdHuman dataset and executes real-time inference of almost 3x∼7x faster than competitive methods.