Research On Pedestrian Detection Algorithm Based On Head-shoulder Under Occlusions

Pedestrian detection,as an important topic in computer vision,has important academic research value and great significance in industrial applications,such as video surveillance and autonomous driving.However,pedestrian detection under occlusions represents a great challenge in real-world applications.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 for full body prediction to boost pedestrian detection accuracy.In this thesis,we proposed a novel bi-branch pedestrian detection network called PedJointNet,that simultaneously regresses two bounding boxes to localize the headshoulder and full body regions based on the unique features of the head-shoulder and full body.Then,instead of using a popular general object detection framework like R-CNN series,SSD,or YOLO,we designed a novel backbone network for pedestrian detection specifically.The one-stage candidate boxes generation method combined with the effective multi-stage feature fusion module,including the pyramid feature idea and the dilated convolution,was used to expand the receptive field of feature map.Moreover,unlike the traditional strategy of keeping the weights fixed for each branch,we designed an inbuilt mechanism during training to adjust the relationships of the head-shoulder and full body predictions dynamically and adaptively,especially under occlusions.We validated the effectiveness of the proposed method using the CUHK-SYSU,TownCentre and CityPersons pedestrian datasets.On the Reasonable and Heavy subsets of the CityPersons dataset,we achieved a miss rate of 13.45% and 52.17%,respectively;on the CUHK-SYSU and TownCentre datasets,we achieved mAP of 78% and 88%,respectively;In terms of detection speed,our PedJointNet also performed well.The single image detection speed on the CityPersons,CUHK-SYSU and TownCentre pedestrian datasets performed best compared to other methods,namely 0.42 s,0.29 s and 0.34 s per image.Overall,compared with other state-of-the-art methods,our two-branch prediction approach has achieved excellent performance in detecting various scales pedestrian,especially under circumstances involving changing occlusions.