Research On Traffic Sign Detection And Recognition In Complex Scenes

At present,with the rapid development of artificial intelligence technology in China,more and more traditional industries have adapted to the needs of the intelligent era,and assisted driving technology and unmanned driving have entered people ’s vision.Traffic sign recognition also plays a vital role in the field of intelligent transportation.However,due to force majeure factors such as illumination changes,complex background information,and tilted shooting angles,traffic sign detection and recognition are more difficult.As one of the most popular algorithms in current target detection technology,YOLOv5 can fully identify complex information in traffic sign images.Therefore,this paper will use YOLOv5 s as the target algorithm to study the complex environment,traffic sign detection and recognition and model performance indicators,so as to improve the ability of model training.The specific work is as follows :Traffic signs in complex scenes are taken as the research object.Firstly,the obtained open source data is processed.After comparison,the CCTSDB data set is selected and the data set is amplified by Open CV,and the traffic signs of complex scenes such as dark natural light,rainy and foggy weather are added.Complete the construction of specific categories,sample balanced fine classification data set GCCTSDB for training.The experimental analysis shows that the new data set has high detection accuracy and good robustness in YOLOv5 s.Secondly,an improved YOLOv5 s model is proposed to solve the problem of high computational complexity due to the deep convolution level in the feature extraction of the original model.Shuffle Net-Block and Mobile Net V3 are used to lightweight the backbone network of YOLOv5 s respectively,forming a comparative experiment.The inverted residual model is introduced,and the traditional convolution of Neck part is changed to deep separable convolution to reduce the complexity of the model.The comparison of experimental results shows that the improved Shuffle Net-Block-YOLOv5 s lightweight model greatly reduces the number of parameters and the model.Finally,the paper uses Qt Designer to design and implement the Graphical User Interface(GUI)for traffic sign detection and recognition,users can log in to the image visualization interface from account information,and then select the optimized model weight according to their needs,and can identify traffic sign information from the collected road pictures and open cameras,which has important reference significance for future research in the field of intelligent transportation.