Research On Pedestrian Detection Algorithms In Dense Scenes Based On Faster R-CNN

Pedestrian detection is a computer vision task which is developed on the basis of the object detection task and is specifically aimed at the pedestrian,and it is usually used to obtain the accurate positions of pedestrians in a single frame of image and a video by using the digital equipment such as computer and camera.At the same time,pedestrian detection is also a front technology of many pedestrian technologies,which is applied in intelligent transportation,safety monitoring,assisted driving of vehicles and many other fields,so it has a very high commercial and research value.However,due to the diversity of pedestrian detection scenes,problems such as human occlusion under dense scenes,multi-scale problems and camera out-of-focus all increase the difficulty of pedestrian detection.Classical object detection algorithms such as Faster R-CNN cannot solve the pedestrian detection problem well,especially in dense scenes.In order to improve the performance of pedestrian detection algorithm and expand image pedestrian detection to video pedestrian detection,this paper analyzes and studies the related algorithms of pedestrian detection,and then proposes the image pedestrian detection algorithm and video pedestrian detection algorithm in dense scenes based on Faster R-CNN.This paper first analyzes the main difficulties of pedestrian detection in dense scenes and proposes a multiscale dense pedestrian detection algorithm based on Faster R-CNN.By introducing the feature pyramid network into the feature extraction network to solve the multi-scale problem,the algorithm effectively enhances the detection rate of small scale pedestrians.To solve the problem of pedestrian occlusion in dense scenes,the algorithm proposes a loss function designed for pedestrian targets that are easily occluded in dense scenes to reduce the influence of pedestrian occlusion on the detection effect.Furthermore,aiming at various problems in video,this paper proposes a video pedestrian detection algorithm integrating multi-target tracking clues on the basic of the multiscale dense pedestrian detection algorithm based on Faster R-CNN.In this algorithm,the multi-target tracking module is introduced to obtain multi-target tracking clues.Then,in order to use multi-target tracking clues to assist pedestrian detection,the results of the multi-target tracking module are correlated with the output data of the pedestrian detection network,so as to obtain the more stable detection boxes.A variety of experiments have proved that compared with Faster R-CNN,the algorithm proposed in this paper has significant improvement in average accuracy and recall rate of pedestrian detection.The algorithm can also ensure high stability in complex scenes in video,so it has a high application value.Moreover,the average accuracy of it on the MOT17 DET dataset reached 0.88,which is competitive with other advanced algorithms.