Research On Pedestrian Detection Method In Dense Scene Based On YOLOv3

In recent years,artificial intelligence technology has developed rapidly and has become a hot spot in the current scientific research and industrial fields.The current era has also begun to gradually enter the era of intelligence.Computer vision is an important branch in the field of artificial intelligence,and its purpose is to allow computers to obtain the same visual understanding capabilities as humans.Pedestrian detection is one of the important applications of computer vision,which is widely used in intelligent construction,such as human-computer interaction,intelligent monitoring,and intelligent assisted driving.Pedestrian detection in actual scenes is quite challenging.Because of the differences in the appearance and posture of pedestrians,mixed backgrounds,occlusions between pedestrians,and many other factors,in order to improve the accuracy of pedestrian detection,this paper launches the following work:1.In order to solve the problem of occlusion,small size and complex background that affect the performance of the algorithm in pedestrian detection,three improvements are made on the YOLOv3 pedestrian detection method.First,for a single pedestrian target detection,the original YOLOv3 multi-classifier is changed to a single classifier,and the output dimension of the network is modified.Second,the pedestrian data in the PASCAL VOC2007 dataset is sorted out,and the K-means++clustering algorithm is used to obtain more accurate Anchor width and height based on the prior information of pedestrian width in the data.Finally,CIo U loss is used to replace the regression loss of bounding box,and freeze training method in migration learning is introduced to train the network,which accelerates the convergence speed of network model training and improves the accuracy of pedestrian target positioning.Finally,the improved YOLOv3 is compared with SSD and Faster R-CNN.The final experiment proves that the improved YOLOv3 can significantly improve the detection performance.2.This paper conducts research on pedestrian detection in specific dense scenes.First,it builds its own pedestrian data set DS-People in dense scenes,and then proposes the SE-YOLOv3 network model of joint attention mechanism and feature fusion,which is based on Based on YOLOv3,it focuses on solving the two shortcomings of YOLOv3 when detecting pedestrians in dense scenes: a large number of label rewriting and unreasonable Anchor distribution.SE-YOLOv3 increases the input image size of the network and adds the attention mechanism SE module to the feature extraction network to weight the features of different channels,emphasizing the importance of features,and uses stepwise upsampling to aggregate volumes through Hypercolumn technology The features extracted by the product network produce a single-scale output with high resolution.The final experiment shows that compared with the original YOLOv3,SE-YOLOv3 has significantly improved pedestrian detection performance in dense scenes on the DS-People data set.Finally,we summarize the above research work,and comb out the research directions that are worth exploring in the future.