Vehicle Behavior Prediction Based On Combined Neural Network

Driving behaviour prediction plays a very important role in autonomous driving and assistant driving technology.The capabilities of excellent driving behaviour prediction can greatly improve the operating efficiency and safety of autonomous vehicles.However,in real scenarios,the complexity and uncertainties of traffic flow bring great challenges to driving behavior prediction,making it harder and the accuracy is low.Deep learning algorithm bases on massive data and powerful learning capabilities.It has the characteristics of simple modeling and good generalization capabilities.A reasonable design of the deep learning network structure can achieve higher accuracy driving behavior predictions.However,deep learning algorithm has the shortcomings of over-reliance on data and lack of interpretable analysis on driving behavior prediction problems.This paper proposes a driving behaviour prediction model that combines three algorithms of Gradient Boosting Tree(GBDT),Convolutional Neural Network(CNN)and Long Short-Term Memory network(LSTM)to fully mine driving behaviour characteristics.The advantages of GBDT in interpretability and the strong generalization ability of CNN-LSTM are used to obtains a high quality prediction result.The main work is as follows:Firstly,this paper uses the CNN-LSTM parallel network model to predict the driving behaviour.In traffic scene,the driving behavior of the target vehicle in the future will be affected not only by its own historical motion information but also by the historical motion information of surrounding vehicles.This information contains the characteristics of two dimensions in time and space.LSTM and CNN are used to extract the spatiotemporal features in historical driving information of vehicles.This paper divides driving behavior into three types: left lane change,straight,and right lane change.Through statistical analysis of different driving behavior samples,the characteristics of different driving behaviors are obtained,this paper uses the historical movement information of the target vehicle and its surrounding vehicles to predict the future driving behavior of the target vehicle.After simulation experiments,the CNN-LSTM model has better prediction performance than traditional behavior prediction models.Secondly,this paper uses the GBDT algorithm based on decision tree to conduct feature mining on the driving behavior rules of the vehicle during driving,which can make up for the lack of interpretability of the CNN-LSTM model.The CNN-LSTM model is essentially an end-to-end structure which cannot describe the interactive relationship between different driving behaviors taken by a driving vehicle and its surrounding vehicles.The GBDT algorithm used in this paper can quantitatively describe the interaction between the vehicle and its surrounding vehicles during the driving process,obtain a series of driving behavior rules when the driving vehicle adopts different driving behaviors,and integrate the driving behavior rule features into the CNN-LSTM combination neural network,the accuracy of the driving behavior prediction model is further improved.After simulation experiments,the decision tree can quantitatively describe the interactive relationship between driving vehicles.At the same time,the driving behavior prediction model that incorporates quantitative features has better prediction performance than the CNN-LSTM model.There are 25 figures,8 tables and 62 references in this paper.