Research On Vehicle Pedestrian Target Detection Algorithm Based On Deep Learning

With the development of science,technology and people’s wealth is increasing,the increasing number of vehicles has also caused frequent traffic accidents,that caused serious harm to people’s lives and property safety.Therefore,the development of automotive intelligent assisted driving technology and unmanned driving technology that has become an important research direction.Among them,the research of deep learning target detection technology based on visible light and infrared images that has great valued,due to the diversification of vehicle and pedestrian targets,the complexity of the road environment makes vehicle and pedestrian detection technology is very challenging.This paper makes a comparative and systematic analysis that has researched on the detection of vehicle pedestrian targets,and proposes a detection framework based on the fusion of infrared and visible light images.The main work is as follows:(1)Studied the basic elements of deep learning,analyzed the important performance indicators,and involved deep learning network model,Then selected FLIR,BDD100 K and KAIST multispectral data sets as the original data sets.Finally analyze and organize them.(2)This paper proposes an image fusion method based on the lightweight MobileNetV2 network.Firstly,used the five layers 1,3,10,16 and 18 of the visible light image detail part and the infrared image detail part that the MobileNetV2 model are merged into a new detail part through the maximum value.Then reconstruct that used the basic part of the visible light and the detailed part that fusion using the weight addition method into a new image,Finally,calculated the FMIdct,FMIw,SSIMa and Nabf’s values that fused image to compare subjectively and objectively with other methods.(3)Based on the analysis that of the one-stage detection algorithm,an improved YOLOv5 multi-scale prediction network that is proposed to detect the target in the images,so that the network has a larger scale prediction range and higher m AP(m AP@0.5 and m AP@0.5:0.95 is respective higher 0.07 and 0.06),whilemean the accuracy rate,recall rate and reasoning speed of the network model that optimized respectively 0.06,0.04 and51 FPS than the original network.(4)Experiments is performed on the infrared,visible light,hybrid and fusion data sets in the FLIR,BDD100 K and KAIST multispectral data sets.The experimental results show that the model generated by the fusion data set training is more accurate,recall and m AP value than the mixed data set training model.High,the model generated by the mixed data set training is better than the model generated by the visible light and infrared data set training.This paper first uses the image fusion method that is the lightweight MobileNetV2 network to fuse the infrared image and the visible light image,and detects the vehicle and pedestrian information on the fused image,then predicts the type,location and probability of the target in the fusion image.To solve the detection problem of vehicles and pedestrians in complex situations.