Research On Pedestrians Detection Method Based On Deep Learning

Pedestrian detection,as a key technology in the field of computer vision,has been applied in many fields,such as human behavior analysis,intelligent transportation,intelligent video surveillance,intelligent robots.In the field of assisted driving,the environment where pedestrians are located is outdoors,and the weather changes frequently outside,making pedestrian detection in traffic scenes still very challenging.The thesis focuses on the research of pedestrian detection algorithms in traffic scenes based on neural networks,with the goal of improving detection accuracy and ensuring detection time.Aiming at the current pedestrian detection algorithms that have many missed and false detections in complex traffic scenes,a method is proposed.The improved regional proposed network is used to detect pedestrians in traffic scenes.(1)Established a data set in the traffic scene,and completed the establishment of the data set through operations such as collection,screening,and preprocessing.This paper selects 1298 images in the training and verification data set and 179 images in the test set from the public data set and the pictures of the actual scene intercepted.The ratio of the training set and the verification set is set to 7:1.The picture should include at least one pedestrian walking upright.(2)Finished cluster analysis of candidate frames in RPN network.The size of the anchor frame in the original RPN network is a general size designed according to the used data set,and cannot be well applied to the detection of a single pedestrian target in a traffic scene.Therefore,the K-means clustering method is used to improve it.The specific method is to read and analyze the tags of all pedestrians,and obtain 9 different anchor frame sizes to make them more in line with the actual ratio.The experimental results show that the AP value before and after clustering has been improved by 9.33%.(3)Researched and experimented the feature extraction backbone network in two aspects: on the one hand,a lighter Mobile Net is used to replace the original VGG convolutional network;on the other hand,partial structures of the two networks are merged,and the merged volume Product network as a new feature extraction network.The experimental results show that the AP value before and after the improvement has been improved by 12.02%.