Research And Improvement Of Pedestrian Detection Algorithm For Occlusion Of Vehicle Scene

Pedestrian detection is a sub-problem of object detection that only locates and recognizes pedestrians.It has always been a research hotspot and difficulty in the field of computer vision.At the same time,it has extremely high application and research value.Pedestrian detection can be effectively combined with tasks such as pedestrian re-identification and pedestrian tracking,applied to intelligent driver assistance systems,intelligent traffic monitoring systems,traffic statistics,and analysis of human behavior in the real scene.The varied postures of the human body,inconsistent imaging methods,different wearing preferences,complex scene image backgrounds,and occlusion between pedestrians and other objects and pedestrians with each other make a great challenge to the efficiency and accuracy of pedestrian detection models.Among them,the occlusion problem most affects the detection performance of the pedestrian detection model.Most of the existing pedestrian detection models based on deep convolutional neural networks make targeted improvement plan based on the general object detection model,and few in-depth explorations of the occlusion problem and effective processing;the complexity of the pedestrian detection model is too high and still does not meet the real-time requirements.A large number of super-parameter settings ser Io Usly affects the detection performance of the model,making the model training and inference extremely difficult.This paper mainly aims at the occlusion problem in the vehicle scene,starting from the two aspects of inter-class occlusion and intra-class occlusion,and draws on the improvement of the existing target detection network idea to make it have more superior pedestrian detection performance.The following aspects are the main work of this article:(1)Drawing on the idea of separate prediction components for human body segmentation,it is proposed to divide the human body into multiple regions,respectively predict var Io Us parts of the human body and score the visibility,and draw on the idea of generating accurate proposals in multiple stages,and design a multi-stage cascade detection network.Experimental results show that this method is more robust than other two-stage pedestrian detection models in terms of occlusion,and it is more accurate in detecting the position of the human body.(2)Drawing on the design ideas of focal loss and repulsion loss,an occlusion perception force loss function is designed to solve the problem of occlusion within a class,and an improved visibility scoring strategy is used to reduce the impact of occlusion among classes.In order to speed up the calculation speed,the prev Io Us two-stage or even multi-stage cascade detection method is abandoned,and the one-stage detection model is adopted,which greatly reduces the detection time.Experimental results show that this method can solve the occlusion problem well and has higher computational efficiency compared with other two-stage and one-stage pedestrian detection models.(3)Drawing on the design ideas of ancho-free object detection,the two types of ancho-free object detection methods are merged,pedestrian detection is redefined as a series of point predictions,and the center point and scale prediction are used to modify the bounding box.Experimental results show that this method has better occlusion detection performance than other pedestrian detection models based on anchor boxes and without anchor boxes.