Research On Pedestrian Detection Methods Based On Deep Learning

Pedestrian detection is an indispensable technology for robotics,smart home systems and smart video surveillance.Mainly refers to the positioning and recognition of the target of the pedestrian in the video or image.Fast and effective pedestrian detection methods play a key role in every aspect of life.However,pedestrians in real life scenes often have characteristics such as insufficient lighting,complex backgrounds,different clothes,dense crowds,high occlusion,and multi-scale,which make the research in the field of pedestrian detection quite challenging.The existing pedestrian detection technology is not effective in detecting pedestrians in real life,and there are many problems to be solved.In view of the above problems,this paper mainly uses deep learning to study pedestrians detection in real scenes.The main research work is as follows:(1)This paper proposes a pedestrian detection method based on Faster R-CNN and multi-layer feature fusion and constructing pyramid hierarchy.In response to the problem that the Faster R-CNN model has poor performance in occlusion pedestrian detection,In order to effectively use the information of the multi-layer convolution output feature map,The feature extraction model of the original Faster R-CNN network architecture is subjected to different convolutional layers for feature fusion,A pyramid-level RPN architecture is designed,and experimental analysis of different improved algorithms is performed to verify the feasibility of the method.(2)Design a pedestrian detection method based on the YOLOv3 algorithm that integrates residual dense units.First,perform cluster analysis based on the pedestrian data set.Second,because the YOLOv3 network has a higher number of layers,in order to reduce the loss of layer-by-layer information,The deep layer of the network introduces residual dense modules,and additionally introduces local and global feature fusion methods to further improve the accuracy and recall rate of pedestrian detection.