Research On Vehicle Behavior Analysis Method Based On Deep Learning

In autonomous driving scenarios,the intention of surrounding vehicles is crucial to improve the decision rationality and control safety of autonomous vehicles.The focus of this thesis is to detect lane-changing and steering intentions of surrounding vehicles(target vehicles)adjacent to an autonomous vehicle.The accuracy of vehicle trajectory prediction is crucial to improving the accuracy of driving intention recognition.Based on this,this thesis proposes a driving intention recognition method combined with vehicle trajectory prediction,and studies the methods in the task of vehicle trajectory prediction and driving intention recognition in dynamic traffic scenes.The main research contents are as follows:Ⅰ.Construct a vehicle behavior dataset based on the classic NGSIM dataset.First,through data preprocessing,the noise jitter and outlier convexity in the original trajectory data are eliminated,and secondary features such as heading angle are added at the same time;local coordinate transformation is introduced to realize the conversion from the road-end perspective coordinate system to the vehicle coordinate system;then describe The method of making behavior tags takes the vehicle’s steering and lane-changing behavior as the research focus of this thesis,and uses the relative positional relationship to extract the interaction features between the target vehicle and neighbor vehicles;finally,extracts the vehicle’s steering and lane-changing trajectory data as data support is provided for training and testing of subsequent methods.Ⅱ.Aiming at the problems that the UB-LSTM model does not consider the time dependence of the vehicle trajectory,exists a large amount of parameters,and ignores the behavior characteristics,a vehicle trajectory prediction method based on the Bi LSTM-GRUAT model is designed.On the basis of the original input features,the behavior characteristics of the vehicle are introduced,and the parallel structure design is adopted.The Bi LSTM branch extracts the forward and backward context information.The GRU-AT branch uses GRU to reduce the calculation amount of updating the hidden state,and introduces the attention mechanism to capture multi-time step information in the time window improves the accuracy of trajectory prediction.The experimental results show that,compared with the UB-LSTM model,the method in this thesis has more advantages in trajectory data modeling,and the accuracy of trajectory prediction is higher.Ⅲ.Aiming at the problem that the GNN-RNN model does not consider the influence of time dependence in the vehicle’s historical trajectory on future trajectory prediction,a vehicle trajectory prediction method based on spatio-temporal graph neural network is designed.Using an encoder-decoder structure,a GAT-based spatial attention module is used to simulate the spatial relationship between the target vehicle and neighbor vehicles,and a temporal attention module based on a multi-head attention mechanism is introduced to capture the relative influence of different historical time steps on future trajectory prediction.Vehicle motion features are extracted using a historical feature encoder.The feature fusion module aggregates the historical motion features,spatial features and time features into a comprehensive feature representation of the target vehicle at each historical time step,enhancing the space-time dependence of the vehicle trajectory,and the trajectory decoder generates the trajectory of the target vehicle at future time steps.The experimental results show that,compared with the GNNRNN model,the vehicle trajectory prediction method based on the spatio-temporal graph neural network in this thesis has a smaller prediction error,and the prediction performance is significantly improved.Ⅳ.Aiming at the problem that the Bi LSTM module in the serial CNN-Bi LSTM model cannot effectively use its temporal feature extraction ability,a driving intention recognition method based on the parallel CNN-Bi LSTM model is designed.First,a constant acceleration motion model is used to post-process the predicted trajectory features,so that the predicted trajectory of the target vehicle is closer to its real trajectory;then,a parallel structure design is adopted,using the CNN branch to extract the spatial features in the trajectory data,and using the Bi LSTM branch to encode The dynamic timing information in the trajectory data,and finally the features extracted by the two branches are aggregated as the input of the fully connected layer,and the final driving intention information is output.The experimental results show that the accuracy of intent recognition of the parallel CNN-Bi LSTM model in this thesis is significantly better than other models.