Research On Deep Learning Based Traffic Signs Detection And Recognition

Traffic signs contain meaningful road information.Autonomous driving system needs to detect them correctly and maneuver accordingly,which is closely related to driving safety.Based on the research of domestic and overseas traffic sign detection technology,this thesis attempts to solve the problems of low recognition accuracy and poor translation scalability of current methods based on deep learning.A domain adaptive traffic sign detection network based on dense high-resolution network is designed.The main work of this thesis is as follows:A feature extraction network with high spatial accuracy and dense connection is proposed.Traditional image classification networks cause the problem of low resolution feature map substantially lose the spatial information.Through the analysis of the high resolution deep network(HRNet),this thesis makes use of the advantage that the feature map produced by HRNet has rich spatial feature information.To refine its low resolution sub-networks training process due to its shallow depth,adding dense connection to all sub-networks.At the same time,all the different resolution feature maps generated from all sub-networks are integrated to make the final feature map has more semantic information.As the feature extraction network of Faster RCNN model,the traffic sign detection result is further improved.A traffic sign domain adaptation method based on Faster RCNN model is proposed.In order to improve the translation scalability of the traffic signs detection network in the scenarios that complex and lack of effective annotation data.The H-divergence is applied to describe the feature distribution distance between the source domain and the target domain.Analyzing the problems of current domain adaption method in traffic sign detection task.Aiming at the different domain shift granularity between traffic signs and entire image,refined the current domain adaption method.Global and the local domain adaptation modules are designed respectively.Improving the model robustness in various test scenarios.Through the experiments on the common traffic sign detection data sets,the feature extraction ability of the dense high resolution network proposed in this paper is better than the traditional image classification network.The proposed local and global module is tested in single dataset and cross domain datasets.The experimental results show that the algorithm in this thesis has achieved more than 80% recognition accuracy on different types of traffic signs.Compared with the original Faster RCNN algorithm,the recognition accuracy of the detection network with domain adaptation module has increased about 20%.Further improved the current domain adaption method.