Road Traffic Sign Detection And Recognition Based On Improved Faster R-CNN And YOLOv3

With the development and update iterations of automobile intelligence and active safety technology,the detection and recognition methods of traffic signs are also in a process of in-depth research and exploration,from traditional color and shape detection methods to deep learning algorithms.This paper takes the car in the road environment as the carrier,based on the Faster R-CNN and YOLOv3 algorithm,to study the working principle of the traffic sign detection algorithm and explore the improvement direction.The main work of the thesis is as follows.(1)Based on the improvement of the Faster R-CNN model,the backbone network VGG16 was adjusted,some convolutional layers and fully connected layers in the network structure were deleted,the number of network parameters was reduced,and the values of some of the parameters were adjusted Fine-tuning;optimizing the regional candidate network RPN,using multi-feature fusion and anchor frame reunion to add a new anchor frame,through multi-scale and multi-task anchor frame,improve the detection accuracy of road traffic signs;proposed improvements The task loss function generation can improve the model’s detection accuracy of multi-scale traffic signs.(2)Based on the improvement of the YOLOv3 model,the first is to adjust part of the loss function in the original model,modify the loss value of the coordinate position prediction in the original function,and add a new anchor box prediction value in the loss function composition.The K-means++ algorithm is used to re-cluster the prior boxes,and the size and number of the prior boxes are adjusted.While adding an a priori box,the average size of the a priori box is reduced,and the target for large scales is improved.The detection accuracy.Adjust the structure of the backbone network Dark Net-53,and add a classification module structure to it to improve the accuracy of detection and recognition.(3)The improved Faster R-CNN model and YOLOv3 model are both tested on the LCTSDB data set,and the improved algorithms are compared and verified in terms of accuracy,recall and accuracy,and finally pass the average The accuracy value m AP reflects the performance of the model.