Attention-based Vehicle Behavior Prediction And Collision Warning

Vehicles are indispensable in people’s daily traveling experience.In recent years,there are more and more traffic accidents caused by vehicles changing lanes randomly,severely affecting the security of people’s life and properties.Meanwhile,with the deepening of studies on artificial intelligence,unmanned driving and aided driving also start to enter people’s life,but security has long remained an important issue faced by us.How to predict the behaviors of other vehicles changing lanes in time and accurately so as to avoid the occurrence of collision accidents is critical to ensuring safe driving.Therefore,this paper studies the issue of predicting and preventing collision by other vehicles to assist in the safe driving of unmanned vehicles.The main content of the research is as follows:This paper proposes an LSTM model predicting vehicular behaviors based on attention mechanism.First of all,it makes use of the Next Generation Simulation data collected by National Highway Traffic Safety Administration of the United States to construct a new data set of vehicle trajectory through the vehicle state characterized by continuous time steps;secondly,it adopts vehicle trajectory data with different step lengths to train LSTM model,and combine the advantages of different models to yield better prediction results than those yielded with single models;finally,attention mechanism is introduced to each time step,so that the model will focus on the data put in at time points containing numerous information in the trajectory data,realizing the real-time prediction of vehicular behaviors and laying a solid foundation for subsequent methods.We also proposes a way of model optimization based on prejudgment of information on the surrounding vehicular environment regarding the over-reliance on the input trajectory length for the effect of model,thereby causing poor real-time performance of the model.The relationship of unmanned vehicles,vehicles on neighborly lanes and other detectable vehicles in terms of geographic location and kinematic snake up the environmental information at the moment.As to environmental features,this paper adopts the training method combining C45 decision-making tree and Bagging integrates learning model so that excellent training effect can be achieved with unbalanced samples.The optimized model can predict 95%of the turnings to the left lane and 61%of the turnings to the right lane.The recalling rate in the prediction result has been improved by around 10%compared to that before optimization.The accuracy of prediction is up to 95%.The effectiveness of the model is testified through the testing data and visualization results.The paper further predicts the driving trajectory of vehicles on the basis of behavioral prediction,and constructs a mathematical model,linear regression model and BP neural network model based on vehicular kinematic information regarding the difficulty in constructing historical trajectory library and weak real-time performance of dynamic generation of trajectories in the existing methods.The models are used to predict the duration of lane-changing respectively.RAN SAC regression method is also adopted to estimate the vertical position of vehicles.In the end,the driving trajectory of vehicles is generated rapidly.On the basis of the prediction of vehicular behaviors,the paper proposes a way of estimating collision likelihood based on the time of collision and the time of relief valve value for changing lanes.First of all,it estimates the duration of changing lanes with linear regression method and calculates the threshold time ensuring the safe driving of vehicles.Then,it calculates the time of collision and sketches a collision probability diagram based on the relative distance,speed and acceleration of unmanned vehicles and vehicles changing lanes at each moment,so as to determine the safety of the driving state of vehicles with the valve value of probability.In the end,it adopts the method of continuous time step to remove the influence of outliers on the judgment result regarding the misjudgment caused by abnormal hopping of acceleration.It also testifies the effectiveness and the method through data testing.