Vehicle Trajectory Feature Recognition And Analysis Based On Deep Learning

Establishing a driver's behavior characteristic model and assessing the driver's accident tendency by vehicle trajectory data mining can bring a wide range of application value to intelligent assisted driving,insurance,driver training and other fields.However,the traditional driver classification method based on trajectory data analysis has some limitations,such as relying on feature engineering based on professional experience,not fully utilizing unlabeled data,and satellite positioning trajectory data cannot characterize specific driving behavior.The development of deep learning models in recent years has opened up new ideas for solving these problems.This paper proposes a trajectory mining method based on deep learning to solve these deficiencies.The main research work is as follows:(1)Aiming at the problem that the feature extraction of current trajectory data depends on artificial experience,which leads to the lack of objectivity and comprehensiveness,a mining method for driving safety characteristics based on convolutional neural network is proposed.This method first preprocesses the original trajectory data,extracts the vehicle motion state data in the trajectory points,and uses it as the input of the network.Then,the convolutional layer is used to automatically extract features,and the fully connected layer is used to classify the extracted features in combination with the sample labels.Finally,the trajectory classification results of the same driver are combined to obtain the safety level of the driver.Experimental results show that the proposed method can effectively distinguish the driving behavior characteristics of different types of drivers,and the accuracy of the classification results reaches 76.4%.(2)In order to solve the problem that the current research fails to make full use of unlabeled data,a semi-supervised network model is proposed based on the convolutional neural network model.The model uses labeled and unlabeled data to train the convolutional deconvolution automatic encoder to obtain the common features in the original data,and uses the labeled data to train the classifier to obtain the information in the label.By sharing the convolutional layer with the auto encoder and the classifier,a unified and streamlined network architecture is obtained.Finally,a two-stage automatic encoder and classifier joint training strategy is proposed to better train the supervised and unsupervised components of the network model.Comparative experiments show that the deep semisupervised learning method proposed in this paper can reduce the number of labeled samples while ensuring the classification performance of the model.In the case of the same number of labeled samples,the classification accuracy of this method is improved to 77.5%(3)Aiming at the problem that the vehicle trajectory based on satellite positioning cannot express the specific driving behavior and lead to a lack of intuitive interpretation of the analysis results,this paper proposes a lane change trajectory analysis method based on multi-channel deep convolutional neural network.Firstly,the key factors affecting the safety of lane change are analyzed,and the trajectories of the lane change vehicle and the other three vehicles are selected to describe the lane change vehicle trajectory sample that describing the lane change process.Then use 3D convolution kernel to extract feature from the sample data and train the convolutional neural network classifier.Experiments on the public data set NGSIM show that the multi-channel deep convolutional neural network lane change trajectory analysis method proposed in this paper can distinguish lane change behaviors with different security,and the classification accuracy rate reaches 82.1%.