Research And Application Of Object Detection And Segmentation Method For Traffic Scene

With the rapid development of autonomous driving technology,new challenges are posed to the real-time and accuracy of algorithms for safety reasons.In this thesis,the target detection and semantic segmentation algorithms based on deep learning in traffic scenes have been deeply and systematically studied.Based on the mature mainstream deep learning algorithms,effective improvement methods have been proposed.The specific work is as follows:1)Aiming at the problem that YOLOv5 is not accurate enough to detect vehicles and pedestrians with small targets in traffic scenes,an improvement method of YOLOv5 network model is proposed.Firstly,the depthwise separable convolution structure is added to YOLOv5,and Mobile Netv3 is used as the feature extraction network,which greatly reduces the parameters of the network and the training cost,and speeds up the detection speed.Secondly,the Leaky Relu activation function in YOLOv5 is replaced by the h-swish activation function to improve the non-linearity of the network model.2)In the Neck part of feature fusion,SPP structure and PANet structure are used to output feature maps of four scales.SPP structure can increase the perception field of feature maps,while PANet structure is used to fuse features of different scales.The new model alleviates the problem of YOLOv5 missing error detection in complex environment,and improves the detection accuracy and speed of the model.3)Aiming at the problem that segmentation image algorithms run slowly and cannot meet the real-time requirements.An improved image segmentation algorithm for traffic scene based on YOLOv5 is proposed.Mobile Netv3 is used as the main feature extraction network.To solve the problem of insufficient spatial perception,atrous spatial pyramid pooling is used to increase the perception field.The experimental results show that the proposed method improves the detection speed and effectively improves the image segmentation effect while guaranteeing the accuracy of the segmentation.4)Developed a traffic scene environment perception system.The system includes the following three modules,namely the scene environment detection module,the user interaction module,and the administrator module.The scene environment detection module is mainly responsible for detecting traffic scenes in the image data uploaded by users.The user interaction module is mainly responsible for implementing some basic operations when the user detects pictures.The administrator module is mainly responsible for the permission management of user accounts.Finally,this thesis combines object detection algorithm with practical application to build an environment awareness system for traffic scenes.