Research On Roadside Traffic Sign Detection Based On Faster R-CNN

Nowadays,the transportation mode based on road traffic has developed rapidly,followed by frequent traffic accidents.In response to this important issue,the intelligent transportation system has gradually developed,and traffic sign detection is one of the key links.However,the traffic sign is very difficult to detect because the actual traffic environment is affected by various external factors.In recent years,with the improvement of computer computing power,it has been widely studied and favored by academia by training deep convolutional neural networks and applying them to target detection methods.Faster R-CNN is one of the typical representatives.This paper mainly studies the traffic sign detection task based on Faster R-CNN.However,using Faster R-CNN for traffic sign detection directly still has the following difficulties: 1)The traffic sign data set is troublesome to obtain,and the manual labeling is time-consuming and laborious.Therefore,there are fewer traffic sign images with labels,and it is difficult to train a good model;2)Because the collected images are interfered by various factors,some images are of poor quality,and there are many small size traffic signs,which poses a great challenge to the traffic sign detection task.In view of the above problems,the main work of this paper is as follows:(1)A data enhancement method based on perspective transformation is proposed,which can make the image produce stereoscopic transformation effect to simulate different shooting angles under real conditions,thus effectively expand the dataset.For traffic sign detection tasks,this method is more efficient than the normal data enhancement method.(2)Due to the small area proportion of most traffic signs in the image area,the number of negative samples obtained by Faster R-CNN’s RPN is far more than that of positive samples,which affects the training effect of the network.Therefore,this paper proposes the method that balance the positive and negative training samples obtained by RPN network,to get better training results.(3)In order to improve the detection effect of poor quality images,this paper proposes to add a threshold segmentation method based on HSV color space before Faster R-CNN,which can remove a large amount of background information,and enhance the saliency of traffic signs in the image,thus reduce the difficulty of further detection.(4)In this paper,the residual network is used as the feature extraction part.With the shortcut connection structure of the residual network,the residual Faster R-CNN with multi-scale connection is constructed,so that the back layer contains more information of the front layer.While obtaining deeper feature expression and there will be no network degradation.And there is a clear advantage for the detection of traffic signs,especially the detection of small size traffic signs.In this paper,the proposed method is compared with the original Faster R-CNN on the acquired data set,and the mAP is improved by 8.6%.Compared with several other effective target detection methods,the superiority of the method proposed in this paper is verified.