Research On Traffic Signal Recognition Method Based On Feature Fusion

Traffic light detection and recognition is one of the key technologies for unmanned and assisted driving,and has broad application prospects.In recent years,it has attracted widespread attention from many scholars at home and abroad.With the continuous improvement of computer computing capabilities,the continuous reduction in the cost of various types of sensors,and the vigorous development of deep learning technologies,more and more studies have been conducted on feature fusion,and they have been widely used in the field of target recognition.The target detection method has extensive application prospect and research value in traffic light recognition.This paper studies the method of traffic light recognition based on feature fusion,and proposes a method of traffic light recognition based on visual saliency and multi-feature fusion based on the characteristics of traffic lights.Explored the traditional method of fusion of manual features and deep learning features.In addition,the end-to-end object detection model SSD was improved and optimized,and a traffic light recognition algorithm based on the improved SSD model was proposed.The idea of feature fusion based on cross-layer connection of the network was improved.The model's ability to detect and recognize traffic lights with a low percentage of pixels.The work of this paper mainly includes:(1)A method of traffic light recognition based on visual saliency and multi feature fusion is proposed.Firstly,the spectral residual model is used to determine the waiting area of the light lamp;then the color,HOG,LBP histogram features of the candidate area are extracted and weighted fusion is carried out;finally,the support vector machine is used to classify and identify the candidate area,remove the background area and identify the status of the traffic light lamp.The experimental results on the data set collected by the vehicle mounted camera show that the algorithm can achieve better recognition effect and basically meet the real-time requirements.(2)On the basis of traditional manual feature fusion,the method of manual feature fusion and deep learning feature fusion is studied.In this paper,the convolution neural network is used to extract the features of CNN network layer of traffic lights.In addition,the Gabor texture features,Canny edge features and LBP texture features of traffic lights are extracted.The neural network is used to fuse the manual features and CNN network layer features to identify the status of traffic lights.Compared with traditional manual feature fusion and SVM classification,CNN can extract more abundant features,and neural network fusion of multiple features has better effect and shorter time.(3)The end-to-end target detection model SSD(Single Shot Multi Box Detector)is improved and optimized,and TR-SSD model is proposed.In this model,the main network of SSD is replaced by the Inception?V3 instead of VGG;according to the characteristics of traffic lights as small targets,the deep features and shallow features of the network are integrated in recognition;the prediction module of the model is modified to three outputs: target location,confidence and light status.The m AP of the improved network model in the LISA data set is 98.6%.