Research On Traffic Sign Recognition And Its Object Detection Depth Learning Algorithm

With the acceleration of urbanization,urban roads become increasingly crowded and traffic accidents occur frequently.Therefore,the intelligent traffic system based on traffic sign recognition has become a research hotspot.Traditional traffic sign recognition methods have some problems,such as complex feature extraction process,simplified data features,easy to be disturbed by the environment,and weak generalization ability of models.In view of the above problems,this paper proposes to use intelligent algorithm to build traffic sign recognition model.This paper using German Traffic Sign Data Set(GTSRB)and Chinese Traffic Sign Detection Data Set(CCTSDB)as carriers,based on deep learning convolution neural network(CNN),studies traffic sign recognition and detection algorithm.Object detection is more complex than object classification,and object recognition is the foundation of object detection,so this paper chooses to study object classification first,and then to further study object detection.Combining with the traditional method,a lightweight CNN-Squeeze model is designed on the basis of one-dimensional convolution.In the construction of deep network,the Inception module is improved to build a multi-channel fusion weighted model SE-inception,which completes the general deep convolution network and significantly improves the recognition accuracy of traffic signs.Traffic sign detection algorithm is to locate the target based on deep convolutional neural network,this paper designs the target detection model of cross convolution,and completed the simulation data on traffic signs.Firstly,In-depth study of CNN network,combining traditional image preprocessing,HOG feature and image clustering method with convolutional network,designing one-dimensional convolution CNN-Squeeze model and applying it to traffic sign pattern recognition.Secondly,Because shallow network can't extract enough image information to complete multi-class classification,this paper fuses residual channel and feature weighted channel according to the characteristics of different convolution models,and selectsactivation function and normalization operation reasonably to optimize network structure.Based on the TensorFlow framework,the multi-channel fusion weighted network SE-inception based on Inception is designed and identified for 43 traffic signs.Finally,As a continuation of recognition,the target detection algorithm pays more attention to the location information of the target.In the TensorFlow framework,a cross-convolution CSSD model based on VGG network is designed to realize the detection of traffic signs in urban roads.For traffic sign images,the CSSD network will frame the location of the logo and give the corresponding category,so that the detection effect can be intuitively felt.