Research On Traffic Signs Detection And Recognition And Tracking And Application Based On Deep Learning

The results of road traffic signs detection and recognition are crucial decision-making criteria for the safe and stable operation of advanced driver assistance systems and driverless systems.However,in the actual application process,the road conditions are complex and there are many unstable factors.Since the road traffic signs are in an outdoor environment,they are easily affected by factors such as lighting conditions,rainy and foggy weather,and are easily disturbed by other similar buildings in the environment.In addition,in the video captured by the vehicle camera,the area occupied by the traffic sign is relatively small,and the vehicle camera is in motion with the vehicle all the time,which will result in the captured video due to the shooting angle and distance,focusing problems and camera shake.There are problems such as ambiguity and deformation in the video,which bring great challenges to the detection of road traffic signs.Therefore,research on traffic signs detection and recognition in real road scenarios is important.This thesis relied on the "Scientists+Engineers" team construction project of Qin Chuangyuan in Shaanxi Province.According to the small range of road traffic signs in the image,complex and diverse types and uneven distribution,as well as the problems encountered in practical applications,this thesis took 72 kinds of traffic signs commonly seen on urban roads in China as the detection objects,proposed an improved YOLOv5s network model and combined improved YOLOv5s network model with the StrongSORT multi-target tracking algorithm to realize a high-precision road traffic sign detection and tracking method while ensuring high real-time performance.The main works of this thesis are as follows:(1)Data set expansion.Aiming at the problems of unbalanced categories existing in the TT100 K data set,this thesis used the data enhancement algorithm to amplify the traffic sign category with a small number of instances in the data set.The original data set was expanded from 6105 images to 14945.This method improved the balance of the distribution of traffic sign instances,made the network model fully learn the target features,and achieved the purpose of improving the detection accuracy of the network model.(2)Improved YOLOv5s network model.Road traffic signs occupy a small area in the image,this will result in low model detection accuracy.In order to solve this problem,this thesis improved the original YOLOv5s network model through four aspects.First,this thesis adopted the improved K-means++ anchor box clustering algorithm re-clusters the data set to obtain more accurate anchors.Secondly,this thesis improved multi-scale feature fusion module for network models to improve the model’s ability to detect small targets.Then a hybrid space pyramid pooling H-SPPF module was proposed to extract richer contextual information.Finally,channel attention mechanisms were introduced to further enhance the model’s ability to extract road traffic sign targets.The model before and after the improvement was compared with the amplified data set in this thesis and the open source data set.The experimental results showed that the precision and the recall rate and mAP@0.5value and mAP@0.5:0.95 value increased by 1.71% and 8.09% and 6.0% and 4.28%respectively on the expanded dataset of this thesis,the effectiveness of improvement strategy proposed in this thesis was verified.(3)The improved YOLOv5s was integrated with the StrongSORT multi-target tracking algorithm.In the real driving environment,the camera will move with the movement of the vehicle all the time,there will be motion blur phenomenon,and there will also be a phenomenon of shaking with the car body shaking.That will result in false detection and missed detection of road traffic signs.For the above problems,this thesis combined the improved YOLOv5s with StrongSORT to detect and track road traffic signs,The results showed that the fusion algorithm could reduce the false detection rate and missed detection rate in video detection while obtaining a stable detection frame in video detection,and the robustness and stability of the detection algorithm were improved.(4)Using the road traffic sign information detected by the improved algorithm,driving safety assistance method was designed.This method included the speed warning safety system and the road blind spot information warning system.Through the comparison of experiments in this thesis,it was shown that:(1)reasonable data set sample equalization processing could help to improve the detection accuracy of the model;(2)By improving the network structure of YOLOv5s,the accuracy of traffic sign detection could be improved without sacrificing detection’s efficiency;(3)The improved YOLOv5s detection algorithm was integrated with the StrongSORT algorithm,which improved the stability and anti-interference ability of the algorithm in video detection and tracking.(4)The proposed driving safety assistance method based on road traffic sign information could provide drivers with a certain degree of driving assistance,thereby it could improved the driving safety performance of vehicles.