Research On Lightweight Technology Of Traffic Sign Recognition Based On YOLOv

Traffic signs play an important role in transportation systems,helping drivers better understand current road conditions,reminding them to obey traffic rules,and playing a role in maintaining traffic order and reducing traffic accidents.Traffic sign recognition,a key technology in autonomous driving technology,has also made significant progress with the development of deep learning.Despite the better performance of deep learning algorithms,their complex models and slow detection speed make it difficult to transplant them to resource-limited in-vehicle hardware devices.Therefore,to address these problems,this paper selects the YOLOv5 s network as the base model and optimizes it,aiming to maintain the detection accuracy while reducing the complexity of the model and improving the detection speed.The details of the study are as follows:First,a lightweight YOLOv5 s algorithm M-YOLOv5 s is proposed to address the problems of the existing deep learning network models that are relatively complex and slow in detection.The algorithm is based on YOLOv5 s by introducing the Bottleneck block in the lightweight network model MobileNetV3 to restructure its backbone network to reduce the complexity of the network model and improve the detection speed.To address the problems of slow convergence and positive and negative sample imbalance in the training process of the network model,the Focal-EIOU loss function is introduced as the localization loss function of YOLOv5 s.Experiments on TT100 K dataset achieved 79.7% recognition accuracy,21.8% reduction in number of parameters,44.3% reduction in computation,and 5.9 FPS improvement in recognition speed compared with YOLOv5 s.Second,an improved M-YOLOv5 s algorithm is proposed to address the problem of excessive accuracy loss due to network lightweighting in the algorithm in Chapter 3.This algorithm is based on the M-YOLOv5 s algorithm and is improved by introducing several optimization strategies.The backbone network is optimized by using the Convolutional Block Attention Module(CBAM)to replace the Squeeze-and-Excitation(SE)attention module in the Bottleneck block,aiming to improve the feature extraction capability of the backbone network.The multi-scale detection is changed from tri-scale to quad-scale,while the large target detection head is removed,aiming to improve the recognition accuracy of traffic signs.Drawing on the structural idea of the weighted Bidirectional Feature Pyramid Network(BiFPN),the Path Aggregation Network(PANet)of M-YOLOv5 s is improved to increase the feature fusion capability of the network,while reducing a small amount of computational cost.The experimental results show that the number of parameters and computational effort do not change much compared with the M-YOLOv5 s algorithm,and the recognition accuracy is improved by 2.9%.Compared with the benchmark algorithm YOLOv5 s,the number of parameters and computation volume are reduced by 21.5% and 41.9%,respectively,and the recognition speed is improved by 5 FPS and the recognition accuracy is 82.6%.Finally,a PyQt5-based traffic sign recognition visualization platform is implemented according to the algorithm in this paper.