Research On Traffic Sign Detection Algorithm Based On Convolutional Neural Network

Traffic sign detection is a key technology in intelligent driving,and it has great research value for unmanned driving and automotive auxiliary systems.In natural traffic scenes,the speed of the car is fast,which requires a higher speed for detection.Compared with other objects on the road,traffic signs are small in size and difficult to detect.This paper uses deep learning technology to study the two important problems of the above-mentioned traffic sign detection,in order to develop a model that can meet the speed and accuracy of the traffic sign detection task at the same time.The specific work content is as follows:(1)Aiming at the difficulty of traffic sign detection and the high precision requirements,this paper improves on the high-precision second-order detector R-FCN model.First of all,the feature extraction network Res Net101 is cropped,and only the first 25 layers are retained,which not only improves the detection speed,but also improves the accuracy.The shape of traffic signs is fixed and regular.The main difference between different types of traffic signs lies in their internal characteristics.With the introduction of deformable convolution and deformable position-sensitive ROI pooling layer,the model can better extract the characteristics of traffic signs.The area of the traffic sign is small,and the original preset anchor point frame of R-FCN is too large.Use the K-means clustering algorithm to obtain an anchor point frame of a suitable size.The OHEM online difficult case mining strategy is used in the training process to reduce simple samples.The experimental results on the GTSDB and CCTSDB data sets show that the accuracy and speed of the improved model have been improved.(2)YOLOv3 is fast and is suitable for detection tasks that require real-time detection speed.For traffic sign detection,this article also made certain improvements on the YOLOv3 model.Replace the original Dark Net53 network with Res Net50 to speed up the detection speed.Deformable convolution is also used in Res Net50,and finally the spatial pooling pyramid structure is used in it.The experimental results verify the effectiveness of the improved model.