Research On Intelligent Detection Method Of Traffic Signs Using Convolution Neural Network

Traffic sign intelligent detection method is the key technology in the field of intelligent transportation,which plays a key role in the process of unmanned driving and high-precision map drawing.In the driverless link,the intelligent detection method of traffic signs quickly and accurately obtains the information of traffic signs in front and transmits it to the subsequent analysis level,which is an important data support for driving path planning,driving behavior decision-making and safety assurance.In the process of high-precision map production,the intelligent detection method of traffic signs quickly and accurately selects the location and category information of road traffic signs from the massive street view image data,greatly reducing the time cost of data processing,which is the necessary link of rapid production of high-precision digital map.In the complex and changeable real road scene,traffic signs will be affected by weather,light changes,shooting angle changes and many other external factors,and there will be changes in color and shape in the image.Similar objects in the road scene will also interfere with the correct detection of traffic signs.How to carry out fast and accurate detection of multiple categories of small-scale traffic signs in the complex real scene is an important problem to be solved in the research of traffic sign detection methods.To solve the above problems mentioned above,this paper studies the intelligent detection method of road traffic signs.Firstly,a regression based traffic sign detection model acb-fpn-net is proposed.The model has a good detection accuracy and speed for small-scale traffic signs in long-distance conditions,and has a good generalization ability in different complex scenes.The improved k-means++ model is used to cluster the anchor frames needed in the process of model detection,and data expansion and data enhancement are carried out on the basis of existing data sets.In view of the limited computing power of the on-board hardware,which can not support the deployment of the general traffic sign detection model in the local and smooth operation,a lightweight traffic sign detection model mg net is proposed.Finally,based on the research of traffic sign intelligent detection method,the platform of traffic sign detection and data management is realized.The main contents of this paper include the following aspects.In order to solve the above problems,this paper studies the intelligent detection method of road traffic signs.Firstly,a regression based traffic sign detection model acb-fpn-net is proposed.The model has a good detection accuracy and speed for small-scale traffic signs in long-distance conditions,and has a good generalization ability in different complex scenes.The improved k-means + + model is used to cluster the anchor frames needed in the process of model detection,and data expansion and data enhancement are carried out on the basis of existing data sets.In view of the limited computing power of the on-board hardware,which can not support the deployment of the general traffic sign detection model in the local and smooth operation,a lightweight traffic sign detection model mg net is proposed.Finally,based on the research of traffic sign intelligent detection method,the platform of traffic sign detection and data management is realized.The main contents of this paper include the following aspects.(2)It is suitable for the research of lightweight traffic sign detection model in low-power scenarios.In order to meet the demand of the traffic sign detection model that can be deployed on the mobile terminal and low-power hardware and run smoothly,the lightweight traffic sign detection model is studied.Aiming at the limited computing power of low-power platform and the demand of traffic sign detection model,a lightweight traffic sign detection model mg net is constructed for low-power platform.Using the depth separable convolution layer to build the feature extraction layer,the network can maintain a certain depth and extraction ability under the premise of limited calculation amount,and using the block convolution to further reduce the calculation amount in the process of feature extraction,and using the channel mixed row to restore the communication ability between the feature channels after grouping,which ensures the effectiveness of feature extraction in the subsequent layers.The maximum pooling and 1x1 convolution kernel are used to form the feature subsampling module,and a lightweight feature extraction network is constructed by overlapping with the feature extraction layer.A double-layer multi-scale detection network is constructed to detect and classify different size targets,and a cross layer connection channel is added to the multi-scale detection network to enhance the reusability of features,and a more semantic feature map is constructed for prediction.The model is tested on a lowpower platform with a low-voltage CPU and a small mobile GPU.The comprehensive performance of the model is evaluated by measuring accuracy,speed,weight of the model after training and comparing with the same type of lightweight detection model.The results show that the detection accuracy of the model on the test set can reach 0.658 map,which is 27% and 12% higher than that of MobileNet-SSD and tinyyolo3 respectively,and it can reach the detection speed of 37 fps on the low-power hardware platform.The size of the parameter model after training is only 21.58 mb,60.2% and 38.1% smaller than mobilenet SSD and tinyyolo3 models,which is suitable for deployment in the hardware condition of low-power and small memory.(3)Design and implementation of traffic sign detection and data management platform.Based on the research of traffic sign detection model,the platform of traffic sign detection and data management is realized.According to different application scenarios,it can achieve real-time road scene data collection,data preprocessing,traffic sign detection based on cloud deployment model,traffic sign detection based on local lightweight model and storage and management of detection results.Through the self-developed mobile Road platform,the functions of the platform are tested in the actual scene,and good results are achieved.