The Research Of Dynamic Multi-scale Pedestrian Detection In Unmanned Driving Scene

Pedestrian detection has a wide range of applications in video surveillance,intelligent transportation,intelligent robots,unmanned driving and other fields,especially in unmanned driving scene,pedestrian detection is an important task related to human life and safety.As an important part of unmanned driving technology,pedestrian detection technology is endowed with strict requirements of high detection accuracy,low delay and real-time performance.However,the detection accuracy of the existing detection algorithms for small and medium-scale pedestrians,especially smallscale pedestrians is not high,which can not meet the needs of unmanned driving scenes.Therefore,this thesis studies the problem of multi-scale pedestrian detection and tracking,in order to improve the accuracy of dynamic multi-scale pedestrian detection and tracking in unmanned driving scenes.Combined with pedestrian characteristics in unmanned driving scenes,a pedestrian detection algorithm based on multi-scale feature fusion is proposed in this thesis.The main research contents are as follows:(1)Design of pedestrian detection network based on multi-scale feature fusion.In view of the difference in aspect ratio and size between pedestrian targets and general targets,this thesis uses PK-means++ method to re-cluster the pedestrian prior frames to make the network more suitable for pedestrian detection tasks;To solve the problem of insufficient learning of small and medium-scale pedestrian features,this thesis proposes a multi-scale feature fusion algorithm MP-YOLOv5 s suitable for pedestrian detection based on YOLOv5 network.The network was redesigned by using the channel attention mechanism,deformable convolutional layers and the improved feature fusion structure.Experiments on City Persons data set shows that the detection accuracy is 94.6%,which is 4% higher than that of the original YOLOv5 s network.(2)Design of dynamic multi-target pedestrian tracking algorithm.In this thesis,Deep-Sort algorithm is used for dynamic pedestrian target tracking.In view of the occlusion and overlap of pedestrian target images,the intersection ratio calculation and cascade matching of Deep-Sort tracking algorithm were improved,and the improved O-Deep-Sort algorithm was obtained.Combined with the pedestrian detection network MP-YOLOV5 s,the complete pedestrian target tracking algorithm design was realized.The improved tracking algorithm can effectively deal with the pedestrian target tracking under occlusion,and the tracking accuracy MOTA is improved to 64.8%,which is 5.2%higher than the Deep-Sort algorithm.(3)Lightweight design of the network.According to the real-time requirements of detection network for unmanned driving scenes,the pedestrian detection network with multi-scale feature fusion proposed in this thesis is light-weighted.Channel pruning was used to model lightweight.Under the premise of ensuring a small change in network accuracy,the detection speed is increased from 106 FPS to 226 FPS,and the model size is reduced by 60.3%,thus meeting the detection requirements of unmanned scenes.Experimental results on public data sets and self-built data sets prove that the pedestrian detection algorithm designed in this paper based on multi-scale feature fusion can accomplish real-time and high-precision pedestrian detection tasks;Pedestrian tracking algorithm based on dynamic multi-objective can realize effective pedestrian tracking.The good performance of the proposed method provides a feasible scheme for the practical application of pedestrian detection in unmanned driving scenes.