Research On Traffic Sign Detection Method Based On Multi-scale Convolutional Neural Network

In recent years,due to the booming of big data algorithms and mobile Internet technologies,intelligent transportation systems and unmanned vehicle technologies have made great progress.Among them,traffic sign recognition is undoubtedly one of the most critical technologies in intelligent driving systems.In this paper,we built a traffic sign detection model based on the convolutional neural network.According to the characteristics of traffic signs,multi-scale features combined with region-based target detectors are used to fuse multi-layer features and train convolutional network models of different structures,as well as using the selected ROIs to train for the location and classification of traffic signs.The work of this paper mainly includes the following points:(1)Particularly,traditional neural network only relies on features of top layer for recognition,which leads to a consequence that spatial information and edge pixels are easily lost.Aiming at the disadvantages,we propose the traffic sign detection method based on region proposal and multi-scale feature.Firstly,we combine the characteristics of traffic signs and feature pyramids to decompose the 7×7 stem block which based on the Res Net network structure into several 3×3 convolutional layers to reduce extraneous parameter and information loss in the original input image;and then,we introduce a dilated convolution in the fourth and fifth residual blocks of the network,trying to expand receptive field without changing the size of the output image.In addition,the high-resolution pixels of the shallow convolutional layer and the high-resolution feature pixels of the deep convolutional layer are fused according to a certain algorithm,and then we design a multi-scale feature network based on region proposals.(2)The training method based on the feature pyramid network model is optimized.The strategies mainly include the introduction of ROI Align algorithm to balance the contradiction between accuracy and detection efficiency.The soft-NMS algorithm which is used for eliminating duplicate windows is designed to better preserve the bounding box of overlapping targets.What's more,we improve the cross entropy loss function to deal with the problem of the limited target pixels and unbalanced distribution of categories.The experimental results show that the optimized traffic sign detection model has different degrees of effect improvement on different traffic sign datasets.(3)In order to solve the problem that part of the traffic signs area in the TT100 K dataset occupies an excessively low proportion of the overall image,we combine the shallow and deep features of the network to build a multi-dimensional feature-fusion traffic sign detection framework.Firstly,enhanced features are generated from the Res Net model to provide basic features for subsequent modules.The backbone of the network consists of codec modules and fusion modules that are alternately stacked.The fusion module combines the maximum output feature map of the previous codec module and the basic features generated by the backbone.Then,the feature maps of the same scale in the multi-level feature group generated in the previous stage are stitched together by the aggregation module,and input to the squeeze-and-excitation module.It is modeled to strengthen the useful feature according to the correlation between the channels.The experimental results on the TT100 K dataset show that the model can solve the problem of low proportion of target pixels,and its detection effect is superior to other models.