Research And Application Of Road Risk Target Detection Algorithm Based On Convolutional Neural Network

With the gradual popularization of the idea of the Internet of Everything,intelligent identification devices have gradually entered our lives.However,whether smart devices can make timely and effective decisions based on the current situation has become a top priority for people.Currently,the more popular method is to use the convolutional neural network method for target recognition.Because the risk targets that may appear on the road have different motion states,different outfits,and different target shooting scenes,light intensity and other factors,the actual amount of data that can be used for convolutional neural network training is relatively small.The recognition accuracy of the convolutional neural network model is inaccuracy.Therefore,in this thesis,aiming at the problems of insufficient data and poor recognition accuracy in road target detection,the convolutional neural network under the limited data set size is studied,and the design of road risk target recognition is more applicable.The convolutional neural network can be reconstructed,and the traditional convolutional neural network can be improved to achieve a new convolutional neural network structure under the premise of insufficient data,which can meet the road risk target that the data volume is far less than the actual requirement identify the use of requirements.The main tasks of this thesis are:(1)Aiming at the conventional combination method in the convolutional neural network model,a re-configurable convolutional neural network model that satisfies the small-scale data set is proposed.The shallower network structure is used to reduce the parameters in the learning process of the re-configurable convolutional neural network using data sets,so that the model learns more common features.In view of the small number of data sets,a data enhancement method is adopted to expand the scale of the data set and improve the recognition accuracy of the re-configurable convolutional neural network.In view of the serious over-fitting of the convolutional neural network,the random inactivation method is adopted to avoid the problem that the convolutional neural network performs better on the training data set and poor performance on the test data set.The experimental results show that the re-configurable convolutional neural network model obtained after training in this thesis can achieve a recognition accuracy of 95% for road risk targets.(2)Aiming at the problem of insufficient training data set size,a VGG-16 network transplantation multiplexing model is proposed.By expanding the original part of the convolutional neural network in VGG-16,the newly expanded network layer is fully connected layer connection for retraining.The VGG-16 convolutional neural network itself has been trained on a large-scale data set.The target features learned by the first few layers of the network structure are more versatile,so only the later network layers with characteristics need to be retrained.That is to say,without adding additional data,it has a large-scale common feature of different targets,and finally a convolutional neural network with an accuracy close to the recognition accuracy on the training set is obtained.This thesis transplants the reused VGG-16 convolutional neural network model,which can realize the accuracy of road risk target identification to 99.5%.(3)This thesis designs and implements a complete prototype system for road risk target detection.For the road risk target detection system,the hardware facilities,software environment and other conditions of its use are analyzed,and the re-configurable convolutional neural network and VGG-16 transplantation multiplexing network proposed in this thesis are implemented.The test of the prototype system under the large-scale data set verifies the feasibility and effectiveness of the model.