Research On Traffic Sign Detection Algorithm_based On Deep Learning

With China’s growing economy,motor vehicles have become one of the basic means of transport essential to national life.The motor vehicles industry has developed rapidly over the past few years as a result of continuous technological advances.As the market demand for motor vehicles continues to expand,the production of motor vehicles is growing annually and the continuous growth of motor vehicles ownership puts the existing traffic system under considerable pressure and challenges,resulting in traffic congestion and a high frequency of traffic accidents.The field of driverless driving is a hot area in recent years,and it can enable existing traffic problems to be solved to some extent.Traffic sign detection encompasses many research directions,and traffic sign detection belongs to a branch of information acquisition in driverless driving.With the current market environment in mind,this article investigates traffic sign detection algorithms based on deep learning and image detection.In order to meet the current market demand for traffic sign detection algorithms,this article investigates and improves the Faster R-CNN algorithm as a basis.Firstly,the basic principle of Faster R-CNN is introduced,and the components of the algorithm are elaborated.Its backbone network is replaced by the Inception Res Net v2 combinatorial network.The experimental part uses the COCO dataset to pre-train the model and fine-tune the parameters,and the German traffic sign dataset is used for evaluation against R-FCN,SSD and YOLOv2 to assess the advantages or disadvantages of each algorithm by various metrics such as accuracy,speed and number of parameters.The experiments show that the improved Faster R-CNN traffic sign detection algorithm has a high detection accuracy but a slow detection speed,while the YOLOv2 algorithm has a high detection speed while ensuring a certain detection accuracy.Based on the above research on traffic sign detection algorithms,and on the advantages of higher accuracy and faster speed of YOLO series algorithms,this article selects the fourth generation algorithm in YOLO series and proposes a traffic sign detection algorithm based on YOLOv4.Firstly,the backbone network of the original algorithm is replaced with Res2Net101,and the loss function is improved by substituting the cross-entropy loss with Focal Loss.In the experiment part,using the Chinese traffic sign dataset,individual traffic signs are randomly selected from the test data set and the PR curves are drawn for analysis.The experimental results demonstrate that the improved YOLOv4 algorithm improves the m AP by 10.4% compared to the original YOLOv4 algorithm,and completes the detection task better in a variety of scenarios,and the improved algorithm has better robustness.A traffic sign detection system has been designed by analysing the actual requirements of traffic signs.The system is embedded with the improved YOLOv4-based traffic sign detection algorithm studied in this article,and contains functions such as image preprocessing,image detection and video detection.The detection system designed in this article is easy to operate and brings some scientific practical value to scholars studying traffic sign detection.