Combination Of Multi-scale Convolution Network And Domain Adaptation Strategy For Traffic Sign Recognition

Traffic signs contain important road traffic information,which plays a guiding role in the driving of vehicles.Timely and accurate perception and understanding of traffic signs information within the vision is the premise to guide intelligent vehicles to run safely and regularly in the road traffic system.Based on the analysis of traffic sign recognition technology at home and abroad,aiming at the problems of low recognition accuracy and poor translation scalability of current algorithms,a traffic sign recognition method combining multi-scale convolution network and domain adaptation strategy is proposed based on the idea of deep learning and migration learning.The main research work of this paper is as follows:A traffic sign recognition network based on multi-scale convolution is proposed.A traffic sign recognition network based on multi-scale convolution is designed to solve the problem that the target size of traffic sign changes greatly and the recognition accuracy is low.Firstly,the basic network is extracted with densely connected features,and the single-scale feature map fusing low-level information and high-level information is extracted from the input image;secondly,the scale conversion layer is constructed,and the different layers of the single-scale feature map are used as input,and the corresponding conversion operation is designed to transform the feature map from single-scale to multi-scale;finally,the target detection block is obtained based on multi-scale feature map.(Candidate regions)and through the prediction network oriented to classification and boundary box regression,the location parameters of boundary box of traffic sign target and its classification probability belonging to each category are output,and the category with the greatest probability is selected as the final classification result.A domain adaption strategy for traffic sign recognition network is proposed.In order to improve the translation extensibility of traffic sign recognition network in changeable test scene which is lacking of effective lable data,a domain adaptive strategy of traffic sign recognition network is designed.Firstly,H divergence is constructed and combined with probability theory to deconstruct the domain adaptation problem,which is transformed into the problem of optimizing the error between image-level and target-level classifiers;secondly,domain classifiers are designed from image-level and target-level respectively,and gradient inversion layer is designed to realize the learning of confrontation;finally,the proposed domain adaptation strategy is applied to traffic sign recognition network based on multi-scale convolution.In the training stage,by inserting a gradient inversion layer between the two feature extraction parts of the recognition network and the domain classifier,the parameters of the traffic sign recognition network and the domain classifier can be synchronously optimized.The domain invariant features can be extracted from the traffic sign recognition network through parameter optimization,and the trained recognition network can be robustly applied to the unlabeled scene data.The proposed algorithm is tested and validated in cross-domain data sets and real traffic scenarios.The experimental results show that the proposed algorithm achieves more than 80% recognition accuracy on different kinds of targets and multi-scale targets.Compared with Faster R-CNN and MC-CNN,the experimental results show that the recognition accuracy of the proposed algorithm is about 30% and 20% higher,respectively.