PedJointNet: Joint Head-Shoulder and Full Body Deep Network for Pedestrian Detection

Pedestrian detection when occlusions exist represents a great challenge in real-world applications, including urban autonomous driving and surveillance systems. However, the head-shoulder feature of pedestrians, which is more stable and less likely to be occluded than other areas of the body, can be used as a complement to full body prediction to boost pedestrian detection accuracy. In this paper, we investigate the unique features of the head-shoulder and full body features belonging to pedestrians. Then, instead of using a popular general object detection framework like R-CNN series, SSD, or YOLO, we propose a novel pedestrian detection network, called PedJointNet, that simultaneously regresses two bounding boxes to localize the head-shoulder and full body regions based on a feasible object detection backbone. Moreover, unlike the traditional strategy of keeping the weights fixed for each attribute, we design an inbuilt mechanism to dynamically and adaptively adjust the relationships of the head-shoulder and full body predictions for more accurate pedestrian localization. We validate the effectiveness of the proposed method using the CUHK-SYSU, TownCentre, and CityPersons datasets. Overall, our two-pronged prediction approach achieves excellent performance in detecting both non-occluded and occluded pedestrians, especially under circumstances involving occlusion, as compared to other state-of-the-art methods.