Research On Traffic Sign Detection Algorithm Based On Deep Learning

The research of autonomous driving technology has been advanced to the stage of open road testing.Traffic signs are used as driving norms,their identification and positioning is a necessary process in environmental perception.Diverse road conditions,harsh weather conditions,unpredictable damage and other factors in realistic scenarios make it difficult for traditional methods to deal with.Vehicle camera sampling methods used commonly result in higher image resolution,and existing deep learning algorithms cannot adapt to the training or inference stage.In view of this status quo,targeted improvements are made after comprehensively considering the problems of insufficient storage,runtime memory,computing power and other aspects of in-vehicle embedded devices.Mainly include the following work:Write a script to count the sample distribution of the commonly used data sets in the field of traffic sign research,select the data set according to the feedback indicators,and use preprocessing methods such as category screening,data division,and sample expansion to clean it.After using the genetic algorithm to improve the K-means method,the data set is traversed to obtain a more accurate anchor box size.For the problem of high resolution,two training strategies of sliding window segmentation and size compression are adopted,and the corresponding post-processing procedures are rewritten.MobileNetv3 and GhostNet are used to replace the backbone of the benchmark network,respectively,and use a very rough hierarchical compression method: discard the most timeconsuming layers of convolution in MobileNetv3 and align the channels.In view of the small proportion of the detection area of the dataset and the decrease in accuracy caused by rough compression,two improvements have been made: the attention module is reconstructed and the spatial attention mechanism based on semantic segmentation is introduced,resulting in an improved algorithm.By comparing with the detection algorithms with excellent performance in recent years in terms of parameter quantity,prediction accuracy,inference speed and other indicators,the comprehensive performance has reached the leading position.A pruning algorithm is adopted to perform adjustable fine-grained compression of the model.By setting the step distance for pruning iteration,the optimal pruning effect is found.After pruning,compared with the original model,the response speed is greatly improved and the loss of accuracy is controlled to a tolerable range.Comparing data with the current mainstream lightweight algorithms,the overall performance remains ahead.The model was deployed in an embedded systems and mobile-oriented CPUs based on Movidius Myriad X chip acceleration and Open VINO framework optimization,the inference time was 294 ms and102ms respectively,which met the design expectations.Finally,the reasoning data of the pruning model in different computing power devices are sorted out.