Research On Traffic Sign Detection Method Based On Improved Yolov5

The detection results and inference time of the traffic sign detection system directly affect the judgment made by the vehicle hub,and ultimately affect the safety of autonomous driving.The detection algorithm based on traditional image processing relies too much on the setting of a priori knowledge,and the robustness is poor,and it is difficult to match the traffic sign detection problem in the complex natural environment.The detection algorithm based on deep learning is highly robust and can adapt to changes in the natural environment,but its performance on small target detection tasks needs to be improved.In addition,the complexity of the model is also one of the factors that limit its deployment on the vehicle side.In order to solve the above problems,this article attempts to improve the existing detection algorithm,selects the current excellent target detection network Yolov5 s as the basic network framework and makes many improvements to it,improving the algorithm’s small target detection performance,and accelerate the inference speed of the model,so it can still maintain high detection accuracy under the lightweight network model.Finally,a traffic sign detection algorithm with high detection accuracy and high inference speed is proposed.The main work of this paper is as follows:(1)Based on the Yolov5 s network model,this paper starts to improve the network structure from the input and output.At the input,increase the size of the image input network,and add a CSP module and one downsampling to the Backbone network to reduce the loss of information when the image input to the network;at the output,this paper adds an output head that focuses on detecting small targets,which greatly improves the detection performance of the model.(2)Using deep separable convolution instead of standard convolution,reducing the amount of model parameters and improving the inference speed of the model;adding SENet to the residual module of Yolov5 s,adding attention mechanism to the network,accelerate the training convergence speed of the model,and improve the accuracy of the model.(3)Use Focal Loss to improve the imbalance of positive and negative samples;use CIoU Loss instead of GIoU Loss to improve the problem that the prediction frame coordinates cannot be effectively optimized in some cases.(4)This article splits the output head of Yolov5 s,and uses three independent branches to output the prediction frame coordinates,confidence and the probability of each category,which improve the detection accuracy of the model.(5)The model is pruned and compressed through Slim Filter Pruner and TensorRT,which greatly reduces the amount of model parameters and the speed of inference.The experimental results show that the improved Yolov5 s traffic sign detection algorithm proposed in this paper has excellent performance: accuracy rate of 92.84%,recall rate of 92.06%,m AP score of 0.936,when deployed on NVIDIA RTX3060 GPU,the inference speed can reach 270 FPS.It has both high detection performance and high inference speed.