Traffic Sign Target Detection Based On Deep Learning

In recent years,convolutional neural network（CNN）has made a breakthrough in the task of computer vision,and has achieved great success in traffic sign classification.In the technology of auto driverless project,traffic sign target detection is a very important task of computer vision.Compared with traffic signs in other countries,China’s traffic signs have their unique characteristics.In this paper,the Chinese Traffic Sign Detection Data Set（CCTSDB）is used as the training object,and the deep learning neural network method is used to study the traffic signs on both sides of the road for target detection.The work done includes the following aspects:（1）Comprehensively learn and understand the development process and technical principles of convolutional neural network,deep residual network and fully connected network.The basic principle of feature extraction network of YOLOv3 algorithm is deeply studied and analyzed.（2）To solve the problem of high undetected rate of small targets,this paper adds a two scale method to detect small targets in traffic signs on the basis of YOLOv3 depth residual convolution neural network architecture model,and improves the prediction regression calculation and loss function of target object frame.（3）In view of the training problem that YOLOv3 needs a large number of data sets,this paper proposes a small sample learning method of transfer learning and twin neural network,which solves the problem of insufficient training samples.The YOOv3++target detection algorithm proposed in this paper,which combines multi-scale feature pyramid and twin neural network,is used to verify on three types of traffic sign data sets,and experimental comparison is made with the method in reference^[62].The experimental results show that the traffic sign target detection algorithm proposed in this paper based on YOLOv3++has an average detection accuracy of 0.989 for all kinds of targets,and the speed of the detected image per second FPS is 45 frames,which is higher than the detection accuracy and speed of the reference,and has achieved good results.