Night Vehicle Target Detection Based On Yolov5s-se And Data Augmentation

With the gradual development of economy,the number of motor vehicles has increased year after year,bringing great convenience to people’s lives.At the same time,the traffic safety situation is severe,with traffic accidents causing serious casualties and economic losses.The night scene is affected by light,and the shape,color and texture of the vehicle are not obvious.Road traffic accidents at night are high,and the accident rate is three times that of the day.Therefore,it is of great practical significance to carry out vehicle detection research in the night scenes to enhance the driver’s perception of vehicle targets.Vehicle detection algorithms have become one of the hot spots in the field of safety assisted driving.The vehicle detection method based on computer vision has the advantages of easy development,low cost,easy transplantation,and strong scalability,so it is widely used.However,there are still the following problems: 1)Vehicle detection is mostly concentrated in daytime scenes,and research on night scenes needs to be improved;2)Vehicle detection based on traditional features is not robust and difficult to adapt to complex scenes;3)Detection methods based on deep learning cannot effectively distinguish the foreground and background regions,and are prone to false detections.Therefore,this paper takes the vehicle target as the research object,based on the principle of deep learning,combined with the yolov5s-se network and data enhancement,to study the vehicle target detection algorithm in the night scene.The main research content of the paper is as follows:1)Design of vehicle detection architecture based on yolov5 network.In the night scene,due to poor lighting conditions,vehicle target features such as geometric features,boundary features,texture features,and symmetry features are weakened,which is not conducive to the development of the thesis.In addition,the existing data sets mainly focus on collecting video sequences in daytime scenes,while there are fewer vehicle target data sets at night.Therefore,the paper first uses the daytime as the research scene,combined with the principle of deep learning,to study the vehicle detection algorithm based on the yolov5 network.Specifically,firstly,compare the typical vehicle target data sets,and preferably the vehicle target data sets that are suitable for the research objectives of this article.Then,through experiments,compare the performance differences of typical target detection network structures,including Faster RCNN,Retina Net,SSD,and yolov5,to explore vehicle detection algorithms in daytime scenes.Finally,in order to enhance the vehicle target detection performance of the algorithm,the attention mechanism is introduced,combined with the SENet network structure,to construct the yolov5s-se network.The experimental results show that the recall rate of the improved yolov5s-se detection network is 1.23% higher than that of the original network,which improves the reliability of vehicle detection.2)Night vehicle target detection algorithm based on yolov5s-se network.On the basis of the yolov5s-se network in the daytime scene,the research on vehicle target detection at night is carried out.First,Cycle GAN is used to generate night-time vehicle simulation pictures.Then,manually mark the collected night real scene pictures,and merge the night simulation pictures to construct the night vehicle detection data set Night Vehicle.Secondly,the yolov5s-se detection network is used to compare and analyze the training results under different data sets.Finally,use the yolov5s-se network to train on the Night Vehicle data set to obtain an NV-yolo network model suitable for night vehicle detection.Experimental results show that the average detection accuracy of the NV-yolo model is 93%.3)Deployment of night vehicle detection algorithm based on embedded platform.The deep learning detection model,due to its huge computing storage capacity,relies on a large computing platform.To solve this problem,the paper develops a visual interface based on Py Qt5,designs a safety-assisted driving night vehicle target detection software,and builds the Jetson Xavier NX platform environment and corresponding hardware Environment,and finally deployed the night vehicle detection network model NV-yolo to the Jetson Xavier NX platform.The test result shows that the average time for detecting a single frame is 46 ms,and the average detection rate is 21 FPS,which basically meets the real-time requirements.