Research On Pedestrian And Vehicle Detection Based On Deep Learning

With the development of science and technology and the increase of people’s wealth,there are more and more vehicles on the road,resulting in frequent traffic accidents,which have a great impact on people’s lives.Road traffic safety problems are increasingly prominent,so intelligent traffic control,the development of autonomous driving technology,advanced assisted driving technology are the hot spots of current research.Pedestrian vehicle detection,as the main body of road target detection,has great research value.However,there are still some problems in pedestrian vehicle detection technology,such as complex traffic environment,different scales of vehicle and pedestrian targets,mutual occlusion,low accuracy and speed of pedestrian and vehicle detection,and missing detection of small targets.In view of this,in order to improve the accuracy and speed of pedestrian vehicle detection,this paper takes attention mechanism and loss function as the focus,and makes a series of improvements to the original YOLOv5s model based on the fast lightweight network YOLOv5s model:(1)Introducing attention mechanics into the YOLOv5s network.The attention machine is introduced into the backbone network of YOLOv5s network,so that the network can extract important features from massive features in the process of feature extraction,weaken useless features and redundant features,and thus improve the detection speed and accuracy of the model.(2)Improve the loss function in YOLOv5s network.SIoU loss function is used to replace CIoU loss function in YOLOv5s model.Because SIoU loss function takes into account the mismatch between the real frame and the predicted frame,such mismatch will lead to slow model convergence and low model detection efficiency,this paper adopts SIoU loss function to improve the detection speed and accuracy of the model.(3)Both the attention mechanism and the improved loss function are introduced into the YOLOv5s model.The attention mechanism is introduced into the backbone network of YOLOv5s model and the loss function CIoU in YOLOv5s model is improved into SIoU loss function,so as to improve the detection effect of the model.Under the same experimental conditions,the experimental results verify that the three improved methods have improved the average accuracy of pedestrian and vehicle detection in a certain degree compared with the original model on the self-made data set mydatas.The third improvement method is the most effective.The AP of pedestrian detection reaches 86.15%,and the AP of vehicle detection reaches 98.23%,respectively increasing by 4.45% and 1.59% compared with the original model.The mAP@.5 of the model reaches 92.19%,3.02% higher than that of the original model,and the model detection speed reaches 71f/s.