Research On The Detection And Recognition Of Traffic Signs Based On The Auxiliary System Of Unmanned Vehicles

Traffic sign detection and recognition plays an important role in the unmanned driving scene.After detecting and recognizing traffic signs,traffic information can be transmitted to the unmanned vehicle and driving decision can be adjusted.In the actual scene,the image captured by unmanned vehicle is usually a high-resolution long-range image,and the proportion of target in the image is small.Aiming at the problems of small area,low detection rate and poor real-time performance of traffic signs in high-resolution images,it is difficult to transplant largevolume model based on deep learning training on embedded platform.The main contents of this paper are as follows:1)In this paper,YOLOv3 algorithm is used to detect and recognize traffic signs.Thirty-three types of traffic signs are selected as detection and recognition targets on Tsinghua-Tencent 100 K data set.The selected data contains noise interference of other unrelated similar categories in the original data,which increases the robustness of the model.2)A multi-scale data set training model of "divide and conquer" is proposed to improve the detection accuracy of traffic signs.In this paper,traffic signs are selected as recognition targets.Through data preprocessing,the image is divided into small resolution data sets and large resolution data sets.Among them,the small resolution data set is to segment the original image in equal proportion,while the large resolution image maintains the original image resolution.In the model training stage,the two data sets based on partitioning are trained on small resolution datasets first,and then finetuned on large datasets through model migration.In the stage of model prediction,the end-to-end method is used to predict traffic signs directly on high-resolution images.Experiments show that the model based on multi-scale data set optimization has a 2.98% improvement in accuracy compared with direct training.3)The model volume was reduced by pruning.Considering the practical engineering application,a pruning theory based on DBSCAN clustering is proposed to prune the model.Compared with the model before compression,the volume of the model after compression is reduced by 4.5%.The experimental results show that the two optimization models proposed in this paper improve the accuracy of traffic sign detection and recognition.Meanwhile,in the platform implementation,the feasibility of traffic sign detection and recognition in engineering application is improved by model pruning.