Research On Traffic Sign Detection And Recognition Based On Deep Learning

With the development of artificial intelligence,self-driving has become the focus of the new phase of research.Because traffic signs can provide the most information while driving on the road,one of the key aspects of self-driving is the driver assistance system,which has the function of detecting and recognizing traffic signs.The difficulty lies in how to quickly and accurately detect and recognize traffic signs.Thanks to the improvement of computer performance and the power of image big data,the performance of deep learning in target detection and object classification ha ve been greatly improved.However,the existing algorithms still have problems such as insufficient real-time performance and low recognition rate due to low detection accuracy.This thesis explores the above issues,the main research contents are as follows:(1)An algorithm for improving the accuracy of target detection frame by combing Ranking Saliency visual saliency algorithm is proposed.The existing YOLO V3 target detection algorithm is used to detect the position of the traffic sign in the image.However,the detection accuracy is low.Therefore,this thesis combines the Ranking Saliency visual saliency algorithm to re-correct the target frame to optimize the target detection accuracy.Experiments show that the combined algorithm increases the IoU of YOLO V3 by 3% ~9% and improves the precision and recall rate of subsequent traffic classifications.(2)In order to further improve the real-time performance and accuracy of the algorithm,a tracking algorithm combined with Kalman filter is proposed.The existing algorithm has the following disadvantages.On the one hand,during the target detection process,frame dropping or occlusion of the target may occur.It is time-consuming to re-identify the target every time the frame is dropped or occluded.On the other hand,continuous recognition has an impact on the real-time nature of the algorithm and reduces the accuracy of algorithm recognition.Therefore,this thesis combines the Kalman filter tracking algorithm to recognize the fixed 10 frames after detecting the traffic sign,and determines the correct category of the traffic sign by voting principle.After the category is determined,the subsequently detected frames are no longer recognized,the algorithm only tracks the classification results.Experiments show that the combined algorithm not only improves the recognition accuracy by 15% and the time performance by 50%,but also better overcomes the occlusion of the target and the frame dropping during the detection process.