Approaches Of Joint Pedestrian And Body Part Detection In Complex Scenes

Joint pedestrian and body part detection based on complex scenes has important theoretical research significance and engineering application value,which is a classic and challenging problem in the field of computer vision.Object detection has achieved great success with the help of convolutional neural networks(CNNs).It can even be applied to industrial fields of great social significance,such as autonomous driving and public safety.Although the object detection method has made significant progress in the past few years,it still remains a challenging problem in joint pedestrian and body part detection on complex scenes because of scale diversity,motion blur,local occlusion,etc.In this paper,we focus on joint pedestrian and body part detection based on the complex scenes.CUHK-SYSU Person Search Dataset consists of many complex scene video images,which are selected to simulate real and complex detection scenes as much as possible.And we research the multi-object detection of pedestrian,head,head shoulder and upper body in the above data sets.We fully investigate the existing object detection methods and propose three detection methods for joint pedestrian and body part detection based on loss function,the relationships among the objects and the training strategy to improve the detection speed(frames per second,FPS)or average precision(AP).Experimental results demonstrate that the proposed strategies outperform the state-of-the-art methods.(1)We novel propose a novel training strategy that use the segmentation loss of pedestrian bounding boxes assisted training model to improve the efficiency of pedestrians and body parts detection.Our strategy introduces a little extra time loss in calculating the segmentation loss of the pedestrian bounding boxes during the training phase.However,in the test phase,no additional time loss is introduced because the branch of the pedestrian bounding boxes segmentation is not required.Based on this,we propose a method that does not decrease the average accuracy(AP)but significantly increases the speed,compared to the state-of-the-art model.(2)We propose a pedestrians and body parts detection strategy based on the body part indexed feature(BPIF)and adaptive joint Non-Maximum Suppression(AJ-NMS),which can improve the average precision(AP).Since some body parts may be partially occluded or even completely occluded,it is difficult to be detected using the information themselves.Therefore,we use the BPIF to encode the semantic relationship between individual components(head,head and shoulders,upper body and whole body)to improve the robustness of detecting occluded components.The AJ-NMS treat one person's head,head-shoulder,upper-body,and body as a whole unit,leading to higher recall for detecting overlapped pedestrians and higher precision for small part such as pedestrian head.Based on the above two strategies,experimental results demonstrate that our method outperforms the state-of-the-art method in joint pedestrian and body part detection.(3)We propose a joint pedestrian and body part detection method based on difficult samples training.We have modified its training strategy based on the state-of-the-art method Head Net.The different positive and negative samples participate at different stages in Bi C and Refinement to fully train the potential of the network.Through our training method,we improve the detection accuracy of pedestrian and body part,especially the head,while the test network is unchanged.