Research On Improved Traffic Sign Recognition Algorithm Based On YOLOv4

Traffic sign detection identification is one of the key technology of intelligent driving system,traffic signs contain a lot of useful information,but it is difficult to detect and recognize traffic signs because of the real environmental impact.Due to the influence of artificial setting factors,the traditional detection algorithms in feature extraction can not meet the accuracy and real-time requirements of detection and recognition of traffic signs under multiple categories.Detection algorithms based on deep learning are favored by researchers because of their advantages of automatic feature extraction and small computation.Based on the current classical YOLOv4 target detection algorithm,this paper proposes several improved strategies,aiming to find a more robust algorithm model to complete the detection and recognition of traffic signs.Specific research contents are as follows:(1)Chinese TT100K(Tsingra-Tencent 100K)data set and American LISA(Laboratory for Intelligent and Safe Automobiles)data set are selected as the training and test data set of traffic sign detection and recognition in this paper.According to the characteristics of traffic sign data set and the shortcomings of YOLOv4 algorithm in traffic sign detection,two feasible improvement strategies are proposed: an improved K-Means clustering algorithm was used to generate anchor boxes,and an improved soft-non-maximum suppression algorithm was used to screen the prediction boxes.Training test and performance evaluation were carried out for each algorithm model under the improved strategy.The experimental results on TT100 K and LISA data sets show that the mean Average Precision(m AP)of the improved YOLOv4 algorithm model proposed in this paper reaches 89.23% and 99.60%,respectively,the recall rate reached 89.26% and 98.80%,respectively,and the detection speed was not affected.(2)In terms of loss function,the SCIo U(Sigmod-CIo U)loss function was proposed by modifying the measurement index Vs on the basis of the CIo U(Complete-Intersection over Union)loss function,in view of the impact of different aspect ratios of the prediction box on the loss results when the real box completely wrapped the prediction box.And based on the SCIo U loss function combined with the above improved strategy,a series of improved strategy combination algorithm models are made.The experimental results show that the improved YOLOv4 algorithm based on the SCIo U loss function proposed in this paper has higher positioning accuracy and better effect than the original YOLOv4 algorithm in target detection.The final algorithm model reaches 89.29% and 99.60% m AP values on TT100 K and LISA data sets,respectively.(3)Aiming at the problems of low detection accuracy and inaccurate positioning accuracy in the current traffic sign recognition task of light weight network,an improved light weight traffic sign recognition algorithm based on YOLOv4-tiny was proposed.The three-scale optimization feature graph strategy and the large-scale optimization feature graph strategy were proposed successively,and the performance of the algorithm model under different combinations of improvement strategies was tested by combining the above improvement strategy.The experimental results show that the improved YOLOv4-tiny model combining clustering anchor box,SCIo U loss function,improved Soft-NMS algorithm and large-scale optimization feature map strategy has the best detection performance.Compared with the original YOLOv4-tiny algorithm model,the m AP and recall on the TT100 K data set are increased by 5.73% and 7.29%,respectively,reaching 52.07% and 64.52%.Compared with the original YOLOv4-tiny algorithm model,the m AP and recall on LISA data set are increased by 2.13% and 5.20%,respectively,reaching 93.48% and 96.47%.In the experimental environment of this paper,the Frames Per Second(FPS)of the improved algorithm model is maintained at about 100 f/s,which meets the requirement of real-time performance.