Study On Two-wheel Vehicle Detection Ahead Of Vehicle Based On Deep Learning

With the development of economy,science and technology,car ownership has increased rapidly,bringing more driving problems and traffic safety problems.Compared with car and car drivers,small-vehicle passengers such as bicyclists still have limited protection on the road and are vulnerable.Most of the residents in rural areas and small towns in China use motorcycles,bicycles,battery cars and other twowheel vehicles,as well as the universal implementation of shared bicycles in large cities in recent years.The two-wheel vehicles have a great impact on traffic safety,so the detection of two-wheeled vehicles is of great significance.But the existing two-wheel vehicle detection algorithm uses the traditional target detection algorithm,which does not achieve accurate detection,and the speed does not meet the real-time requirements.This paper proposes a vehicle-based two-wheel vehicle detection method based on deep learning.The main work is as follows:First chapter analyzes the relevant domestic and international research status of target detection from three aspects: target detection technology,two-wheel vehicle detection and database.It summarizes the three main technical difficulties of two-wheel vehicle target detection,establishes the main research contents and general ideas,and builds Technology roadmap.The second chapter expounds the concept of deep learning and the basic structure of convolutional neural networks.The mainstream algorithms of two-stage detection algorithms based on deep learning such as RCNN,SPP-Net,Fast RCNN,Faster RCNN and the mainstream algorithm of one-stage detection algorithms based on deep learning such as YOLOv1/v2/v3 and SSD algorithm are analyzed and compared one by one.According to the characteristics of the algorithm,the Faster RCNN and YOLOv3 algorithms are used to detect the two-wheel vehicle,and the accuracy,recall,mAP and recognition efficiency are used as model evaluation indicators.The third chapter analyzes and summarizes the data annotation format of the existing open source datasets containing the data of the two-wheel vehicle under the road conditions,and uses the script to clean the existing datasets for different data annotation formats,and the extracted data labels of two-wheeled vehicles are integrated into a unified and trainable label format.And the data samples in the existing data sets are unbalanced,and include mostly foreign traffic scenes as well as the sample size of the two-wheel vehicle is insufficient.The data is collected by the actual vehicle and marked with special data.So The actual vehicle collection data is supplemented and the data labeling is performed using the special data labeling software labelImage,and the existing database is integrated with the actual vehicle collection data into the training data set of the two-wheel vehicle.The fourth chapter builds the algorithm test platform,uses the Faster RCNN algorithm and the YOLOv3 algorithm to train the network model based on the twowheel vehicles data set,and then uses the model evaluation index to compare the two network detection effects.The test results show that the YOLOv3 algorithm has higher detection accuracy.The detection speed is faster and more in line with the needs of twowheel vehicles detection.The fifth chapter analyzes the shortcomings of the YOLOv3 algorithm,and proposes a loss function optimization scheme using the focus loss function instead of the cross entropy loss function and an optimization method for recalculating the size of the anchors using the K-means++ algorithm,and the video offline test is improved.The YOLOv3_two-wheel model algorithm has an accuracy of 92.4%,which is 5.5 %higher than the original model.In the sixth chapter,the research contents of this paper is summarized,and the deficiencies is pointed out,so as to the improvement direction for further research is provided.