Real-time Traffic Sign Detection Research Base On Convolutional Neural Networks

Traffic signs are closely related to automobile,passenger and road safety,so people put forward high requirements for automatic detection and recognition of traffic signs.In addition,improving the accuracy of automatic traffic sign recognition also helps to improve the efficiency of road transportation and create a more efficient and environmental protection transportation system.Convolutional neural network has achieved great success in traffic sign detection.However,sometimes the captured traffic signs are small,which will increase the difficulty of detection.The damage and fading of traffic signs in the wind and rain all year round,the influence of similar distractors in the detection process,and the deformation and fuzziness of the captured picture caused by the high-speed movement of vehicles all make the accuracy of neural network detection of traffic signs low.Some traffic sign detection algorithms are not real-time enough to be applied to real life.The main work of this paper is as follows:In order to detect traffic signs more accurately,thesis propose a detection algorithm based on windowed self-attention and channel spatial attention,which is based on YOLOv5 detection framework.Thesis use the self-attention mechanism to exchange information between each pixel and other pixels to obtain rich semantic information for later detection.In order to slow down the huge amount of computation brought by the self-attention mechanism,Thesis adopt windowing for the self-attention mechanism.In addition,in order to make the information exchange of pixels in the window more efficient,thesis shift the window,but in order to reduce the computational loss,in the experiment,thesis use the scheme of cyclic shift to realize the window shift.In order to further improve the detection accuracy,thesis use the attention mechanism in the two directions of channel and space to emphasize or suppress some feature information.At the same time,thesis improve the spatial pooling layer in the backbone network,extract the most important features in the region with three pooling cores of different sizes,and add three spatial pooling layer modules.In the experiment,the detection speed of the algorithm on GTSDB data set and CCTSDB 2021 data set is 56.82 frames and 56.18 frames respectively,with good real-time performance.The mAP result of the experiment on GTSDB data set is 98.2%,the initial algorithm result is 95.5%,and the mAP result of the experiment on CCTSDB 2021 data set is 84.6%,the initial algorithm result is 95.5%,which proves the effectiveness of the algorithm.A good data set is very important for the evaluation of the algorithm.However,the characteristics and classification of traffic signs in different countries are different,and the amount of data in the existing traffic sign detection data set in China is small.Therefore,our laboratory produced CCTSDB in 2017.In order to solve the problems of incomplete annotation information,no special test set,insufficient attribute classification and the need to expand the number of samples in this data set,we made CCTSDB 2021 data set.The data set removes the images without labeling information,and more than 4000 images are newly labeled.A special test set is also made.The test set has a total of 2000 images,including 500 negative samples,and the rest are positive samples classified according to meaning,weather and lighting conditions and the scale and size of traffic signs.At the same time,thesis selected nine classical objection detection algorithms for the new data set and used six evaluation metrics for all-round evaluation,which provided convenience for follow-up researchers.