Comparative Research On Improved Algorithms Of Traffic Sign Detection And Recognition In Complex Environment

With the increase in the number of cars in my country,it has also caused problems such as traffic congestion and frequent traffic accidents.Intelligent Transportation System(ITS)is a real-time,accurate and efficient traffic management system.One of the important components is the traffic sign recognition system(TSR),The system acquires images on the road surface,detects and recognizes traffic signs on the road,and finally gives the driver corresponding feedback based on the results.At the same time,with the development of intelligent machinery,the traffic sign recognition system is also a very important part of intelligent driving.Based on the target detection network YOLOv4 and YOLOv4-Tiny,this paper does the following work for the detection and recognition of traffic signs.(1)Make a data set.In this paper,the data set is composed of part of the CCTSDB data set and self-collected data set,some data sets are re-labeled with labeling tools,the target tag types are increased,and the data enhancement is completed for some data sets.(2)A comparative study based on the improved YOLOv4 traffic sign detection algorithm.First,compare the model prediction results under different data set enhancement strategies,and the results show that the method of using data offline enhancement and balance is the best.Secondly,conduct a comparative experiment on the three training techniques,respectively are Mosaic data enhancement,Label Smoothing and Learning rate cosine decay.The results show that Mosaic data enhancement and Learning rate cosine annealing attenuation contribute to the improvement of network model performance.Finally,the main structure of YOLOv4 was improved,the attention module and the ASFF(Adaptive Spatial Feature Fusion)module were added,and the enhanced feature extraction network PANet(Path Aggregation Network)was improved,more feature layer inputs were added.The improved YOLOv4 has the best m AP(mean Average Precision),which is 5.98% higher than the original YOLOv4 network.(3)Based on the improved YOLOv4-Tiny’s comparative study on small target traffic sign detection algorithms,in order to improve the lightweight network’s ability to detect small target traffic signs,the following improvements have been made to the network.First,modify the FPN to strengthen the feature extraction network structure,and the lower-level network output features are used,and the more refined feature layer prediction results are used.Secondly,to improve the optimization ability of the model,the Ada Bound optimizer is used to replace the Adam optimizer,and finally the Efficcient Net feature extraction network is used to replace the feature extraction network in YOLOv4-Tiny to obtain stronger feature extraction ability.The result shows that the m AP of the improved YOLOv4-Tiny network is 6.76% higher than that of the original YOLOv4-Tiny network.