Research On Traffic Sign Detection Algorithm Based On Deep Learning

As a key part of automatic driving,traffic sign detection plays an important role in ensuring driving safety.With the rapid development of deep learning,it has become a trend to use target detection algorithm in the field of deep learning to detect traffic signs.However,due to the small proportion of traffic sign targets in the images collected by vehicle cameras,the detection accuracy is low and cannot meet the real-time requirements.Therefore,in this thesis,an improved YOLOv4 network structure based on the representative YOLOv4 algorithm is proposed for traffic sign detection.The main work of this thesis is as follows:（1）Aiming at the problem of missing and misdetection in YOLOv4 network when detecting small target traffic signs,a feature enhancement module is designed to convolve the side output of feature extraction network,and merge the multi-channel features,thereby broadening the feature receptive field and improving the feature richness and expression ability.（2）Aiming at the problem of low precision in YOLOv4 network when detecting small target traffic signs,a target detection layer with more shallow feature information is constructed.The 8-fold down-sampled feature map of the original image is twice up-sampled,and the up-sampled feature map is spliced with the corresponding scale feature map extracted from the backbone network to obtain the scale feature that integrates more detail location information and deep semantic information.And a target detection layer with smaller receptive field is established on the feature,thereby improving the precision of the network in detecting small target traffic signs.（3）Aiming at the problem of slow detection speed in YOLOv4 network when detecting small target traffic signs,a new backbone network DSCSPDarknet53 is constructed.Based on the original backbone network of YOLOv4,DSCSPDarknet53 replaces the standard convolution with the deep separable convolution,thereby reducing the number of parameters of the network model and improving the detection speed of YOLOv4 when detecting small target traffic signs.（4）Aiming at the problem that YOLOv4 network is not effective in detecting occluding traffic signs,the non-maximum suppression algorithm of the network is optimized.The optimized non-maximum suppression algorithm will not directly delete the detection box with a high degree of overlap with the reference box,but attenuate its confidence,thereby improving the network’s ability to detect occluded traffic signs.（5）Aiming at the problem that the default prior frame parameters of YOLOv4 network are not suitable for traffic sign detection task,k-means clustering algorithm is used to optimize the network prior box parameters to make the network have certain pertinency in traffic sign detection and further improve the detection accuracy.Experimental results show that the improved YOLOv4 network can effectively improve the detection ability of small target traffic signs,the mean average precision on TT100K data set is increased by 4.55%,and the detection time of single image is increased by 0.033s.