Research On Technology Of Traffic Signs Detection For Automatic Driving Scene

Environmental perception is the important data foundation of automatic driving technology,involving the comprehensive use of computer vision and automatic control technologies.The ability of automatic driving equipment to perceive the road environment and make control decisions depends on the accurate detection of road elements,which is related to the safety and efficiency of vehicle driving.This article discusses an important type of road elements in depth,namely the detection of traffic signs.In order to make real more accurate and efficient traffic signs detection technology,this paper makes improvements based on Faster Region-based Convolutional Neural Networks(Faster R-CNN),and proposes a deep-learning based traffic signs detection method.Because traffic signs have unique visual characteristics that are different from other road elements,detection algorithms also face different challenges:traffic signs usually have smaller visible areas,and have real-time requirements for algorithms in practical applications.In the series of Region-based Convolutional Neural Network algorithms,Faster R-CNN has achieved good results in solving target detection problems.However,due to the region candidate box in Faster R-CNN,the quality of regional candidate boxes generated by the region-proposal networks are not ideal,and excessive quantity affects running speed,and the network performs poorly in face of small targets.Therefore,it affects the target detection performance of Faster R-CNN,posing a challenge to the detection of traffic signs in its practical applications.Based on this,this paper analyzes the visual characteristics and detection requirements of traffic signs in automatic driving scenarios,as well as the limitations of the Faster R-CNN algorithm to put forward a four-step improvement strategy.First,in order to enhance the effect of the candidate box,the multi-layer convolutional layer features were fused to increase the feature sampling ability of the convolution layer,and the merge pooling method was used in to improve the pooling layer capacity.Second,based on the visual characteristics of small targets,a strategy was proposed to locate the small target through the addition of local context information around the target.Third,the sofmax classification layer and the bounding box regression layer were merged into a single convolutional layer,thereby reducing the size of the output layer,and optimizing the training and testing speed of the network.Finally,the model was iterated and pruned to remove redundant connections,thereby improving the computing efficiency.The experiment in this paper is developed based on TensorFlow.The open source data set Tsinghua-Tencent 100K was chosen to evaluate the detection performance of the method in this paper.The comparison experiment results of this paper’s methodand Faster R-CNN on the data set verified the effectiveness of the proposed improvement strategies.