Research On Visual Detection Algorithm Of Roadside Vehicle Target Based On Deep Learning

Vehicle object detection in roadside perception,as a part of target detection,is not only a key technology in vehicle-road collaborative perception,but also a research hotspot in the field of intelligent transportation.The vehicle target detection system is required to have very low delay and very fast response speed in the roadside perception of vehicle-road collaborative system.The detection model should be as small as possible,and the complex convolutioanal neural network model with excellent performance is not suitable for application in the vehicle-road collaborative perception system.Therefore,it is very necessary to research efficient and sophisticated model for vehicle object detection in roadside perception.Aiming at the application requirements of roadside perception in the field of intelligent transportation,this dissertation studies and designs a roadside vehicle target visual detection algorithm based on deep learning.The main work is as follows:1.Firstly,this dissertation analyzes the current popular object detection algorithms,and compares the two-stage detection algorithm and one-stage detection algorithm,R-CNN series and YOLO series of object detection algorithms.According to the roadside perception system performance requirements of embedded computing platform edge perception algorithm,this dissertation designs a architecture of roadside perception object detection system.2.Aiming at the real-time problem of vehicle target detection in the roadside perception system,based on the lightweight design idea of Mobile Net v3 and the multiscale feature fusion detection structure of YOLOv4,The depth separable convolution,linear bottleneck inverted residual structure and SE lightweight attention module are introduced into the backbone feature extraction network,and a lightweight roadside vehicle target detection algorithm is designed.3.Aiming at the problem that the detection effect of small targets in the algorithm is not ideal,the design of the lightweight roadside vehicle target visual detection network is optimized from the aspects of training strategy and network structure.The training strategy uses the improved Mosaic data enhancement method and the improved regression box loss function to increase the sensitivity of the model to small-scale targets.In terms of network structure,an improved multi-scale fusion prediction structure and SPP spatial pyramid pooling are introduced to enhance the network’s feature extraction capabilities and multitarget multi-scale detection capabilities.4.The dissertation uses UA-DETRAC to train the designed network,and transplant the model to the roadside embedded platform for testing,and finally achieve 92.7% vehicle recognition rate and 35 FPS detection speed on the roadside embedded platform,which proves that the effectiveness of applying lightweight convolutional neural network to roadside vehicle detection.