Traffic Sign Recognition Based On Deep Learning And ELM

In recent years,with the rapid development of domestic economy,people’s demand for cars is increasing year by year,and the incidence of traffic accidents and road congestion is also increasing.Intelligent transportation system(ITS)can alleviate traffic congestion or reduce the frequency of traffic accidents to some extent.Traffic sign recognition is one of the important technologies in ITS and automatic driving application scenarios.Traffic sign recognition system(TSR)can effectively identify traffic signs,improve driving safety and reduce the incidence of traffic accidents.In the natural environment where cars are driving,the road traffic conditions are complex and changeable,which brings great challenges to the detection and recognition of traffic signs.Therefore,it is of great significance to study the traffic sign recognition system with high precision and real time.This paper proposes a traffic sign detection and classification method based on deep learning and extreme learning machine.The research work of this paper mainly includes the following two points:(1)BCS algorithm is proposed in the detection stage of traffic signs.Spatial attention is added to the trunk network of feature extraction to enhance the ability of the trunk network to extract spatial features and guide the model to focus the learning attention on the region where the target may exist.The experimental results show that the BCS algorithm has high detection accuracy and little effect on training and reasoning speed,and has good robustness.The training and testing experiments on TT-100 K data set show that the m AP(mean Average Precision)of the proposed BCS algorithm is 74.1%,which is 15.8% higher than that of the Faster R-CNN network.(2)In the classification stage of traffic signs,a method combining deep learning technology and extreme learning machine is proposed to classify traffic signs.The improved Extreme Learning Machine is used as the traffic sign classifier,and the full connection layer is used to accurately locate the traffic signs,so as to improve the classification accuracy and positioning accuracy.The accuracy of TT-100 K traffic sign data set reached 66.0%,which was 7.7%higher than that of Softmax classifier Faster R-CNN network.The recognition time of each image in the test was 59.13 ms.The experimental results showed that this network had high accuracy and good real-time performance.