Research On Traffic Signal Detection Method Based On Deep Learning

With the continuous development of society,traditional car control methods can no longer meet the needs of the people.Systems such as assisted driving and unmanned driving have emerged as the times require.What follows is the continuous improvement of people’s requirements for the safety of smart car systems.The detection of traffic signs and signal lights outside the cab is very important.Traditional traffic signal detection algorithms are slow,low in accuracy,and poor in robustness.Therefore,it is extremely important to achieve accurate detection of traffic signs and traffic lights by improving traditional algorithms.In order to realize the accurate detection of two types of traffic signal objects,traffic signs and traffic lights,the target detection algorithm YOLOv4 based on deep learning is proposed and improved to obtain YOLOv4-s and YOLOv4-m.In the YOLOv4-s algorithm,the 3 scales of the feature detection network are improved to 4scale detections.First,based on the CSPDarknet53 convolutional neural network,the104×104 feature map output by the third deep residual structure in the backbone feature extraction network The 4th feature detection scale is obtained by fusion processing with the feature map after 3 times of upsampling,and a new feature pyramid fusion network structure is formed,which improves the detection accuracy of small targets and multitarget objects and reduces the missed detection rate.Secondly,the CIo U coordinate positioning loss function is introduced to make the prediction box and the real box converge faster in the regression process,and the calculation of the loss function is more accurate.Finally,the k-means++ clustering algorithm combined with certain a priori box allocation strategy and other important parameters are used in the priori box clustering to conduct comparative experiments to optimize the corresponding parameters and structure of the model.This paper improves on the feature extraction network and convolution method of YOLOv4-s,and proposes the YOLOv4-m algorithm,which replaces the convolutional neural network in the YOLOv4-s feature network structure with a deep separable convolution,and uses the Mobile Net V3 network structure To perform model training for feature map extraction.On the premise that the detection accuracy of the overall algorithm is not reduced,the parameter amount of the model is greatly reduced,and the detection efficiency is improved.In this paper,experiments are carried out on traffic signs,traffic lights and selfmade traffic signal datasets.The experimental results show that the YOLOv4-s algorithm has a recall rate of 87.36% and an average accuracy of 98.12% on the traffic sign dataset LISA;traffic signal datasets On BDD100 K,the recall rate is 77.72%,and the average accuracy is 84.95%.The traffic lights in the complex weather environment in the BDD100 K data set also have better detection results,which are more accurate and real-time than other existing target detection models.And robustness;on the selfmade traffic signal data set TSL,the average accuracy of the traffic signal is 87.25%,which improves the accuracy of small targets while still enabling the detection speed to reach more than 40 frames per second,but the volume of parameters is still huge,YOLOv4-On the self-made TSL data set of the m algorithm,the average accuracy of the second-class traffic signal target is 86.29%.Under the premise of ensuring the accuracy,the detection speed still reaches 51 frames/s,and the parameter amount is only 1/6 of YOLOv4-s.The improved algorithms in this paper have significantly improved the detection accuracy,and achieved the purpose of simplifying the model parameters.