Research On Traffic Sign Detection And Recognition Method Based On Machine Learning

In today’s intelligent driving system,the detection and recognition of traffic signs plays an crucial role.The traditional detection and recognition algorithm has low accuracy and poor speed.The traffic sign detection and recognition on account of deep learning greatly improves its accuracy and speed.But this is only when the objects is large,when the traffic signs appear as a small objects,the detection accuracy is greatly reduced.However,in real driving scenes,traffic signs often appear in pictures in the form of small objects.This research work is devoted to make the detection and recognition accuracy rate of small target traffic signs higher.Firstly,a data preprocessing means is used to reconstruct the resolution of small target traffic signs,in the task of the data set is given priority to with small object traffic signs,made of tsinghua joint tencent TT100 k data sets,it comes from Tencent map street view pictures,picture background and lighting conditions rich,the data sample of traffic signs belong to small goal,but in the sample data category serious imbalance.Before training,sample data are screened and processed to obtain relatively balanced annotation categories.Considering that the proportion of traffic signs in the whole picture is too small and the image resolution is low,the super-resolution reconstruction of traffic signs is carried out to enrich the information in the picture.The classic super-resolution reconstruction model SRCNN is improved,and modified the network structure and improve the activation function to enhance the reconstruction effect of the model.Secondly,the traffic sign detection and recognition algorithm on account of blended attention mechanism and multi-scale features is put frowarded.In this thesis,the network of small object traffic signs is improved based on YOLOv4,it can effectively increase the precision rate of detection and identification,mainly from three parts.Small object itself contains limited feature information,and little effective information is retained after multilayer convolution.It could obtain more feature information by attention mechanism embedded in the backbone network to retain more important information.In the data samples used in the experiment,the percentage of negative samples is much larger than the percentage of positive samples.Focal Loss function was used to ameliorate this disproportion,reduce the influence of negative samples on training model optimization.When extracting image features,the shallow network extracts fine-grained information,mostly geometric features,while the deep network extracts more abstract semantic information,semantic information and geometry information helps to obtain higher test accuracy,the network integrates features of multiple scales,it can obtain more abundant semantic information and spatial geometry information,excellent performance in small target traffic signs detection and recognition,it can well increase the capability of the entire network model.This thesis aims at the means discussed above,the traffic sign data preprocessing,the data amplification,and improve the resolution of the small object traffic signs,the YOLOv4 model was optimized and the performance of each network model is verified on traffic sign data set TT100 k.The results show the method that proposed is resultful,it could have higher accuracy in detection and recognition.