Research On Motion Prediction Methods Of The Target Vehicle

In recent years,due to the continuous improvement of the traffic awareness of Chinese residents,people are increasingly concerned about issues such as traffic safety and traffic efficiency.With the development of autonomous driving technology,the above-mentioned problems will be effectively solved.In an autonomous driving system,accurately predicting the motion of surrounding vehicles will help the host vehicle make better decisions,thereby improving the safety and efficiency.The main work of this article consists of two parts.1)Construct the IC-LSTM model to identify the maneuver of the target vehicle.There are three following characteristics about the recognition of vehicle’s maneuver:The vehicle’s movement is time-sequential;Vehicles’ interaction;The vehicle’s maneuvers are dependent between different time steps(the current maneuver depends on last time step’s maneuver to a certain extent).In order to take the above properties into consideration,this paper proposes the IC-LSTM maneuver recognition model,which contains three modules: input module,interaction module and maneuver recognition module.The input module extracts the information needed by the IC-LSTM model;The interaction module uses utility functions to model the interaction between vehicles;The maneuver recognition module outputs the recognized driving maneuver.The results on the NGSIM-US101 data set show that the accuracy,macro F1 score and the cross entropy in test set is 0.9164,0.8746 and 0.1683 respectively,which is better than the support vector machine model,hidden Markov model and LSTM model.In addition,the model’s average advance time of lane-changing recognition for LLC and RLC is 3.08 s and 2.33 s,respectively,which is better than the comparison models and can provide sufficient redundant time for the decision-making of the host vehicle.Finally,by performing ablation experiments on IC-LSTM,the interaction module,attention mechanism,and CRF introduced in this model contribute to the accuracy rate of 0.0132,0.0043,and 0.0110,respectively,verifying the effectiveness of the relevant modules.In the case of scene verification,the IC-LSTM model proposed in this paper outperforms the comparison model in terms of accuracy,recognition stability,and advance time to recognize lane change maneuver.2)Constructed an I3AP model for predicting the future trajectory of the target vehicle.The predicted trajectory of the vehicle is related to the physical states(position,speed,acceleration,etc.)of the vehicle at the current moment,historical trajectory,driving maneuvers,and the interaction between the target vehicle and its surrounding environment(surrounding vehicles,road structure,etc.).This paper constructs an I3AP trajectory prediction model,which can comprehensively model the above information.The model includes 5 modules: input module,interaction module,maneuver recognition module,attention module and trajectory prediction module.The input module extracts the input information of the I3AP model;The interaction module constructs the spatial interaction feature,historical interaction feature,and future interaction feature successively,which can model the spatiotemporal interaction between the target vehicle and its surrounding vehicles;The maneuver recognition module recognizes the driving maneuver of the target vehicle at the current moment,which is input into the downstream module as the intention feature;The attention module calculates the importance of the aforementioned features through the additive attention model,and weights them by the importance to obtain the final coding feature;The trajectory prediction module outputs the predicted trajectory of the target vehicle via the LSTM decoder.The results on the NGSIM-US101 data set show that the horizontal ADE and FDE of the I3AP model proposed in this paper is 0.1624 m and0.3045 m,respectively,and the vertical ADE and FDE is 1.4512 m and 4.0443 m,respectively,which are better than those comparison methods: model based on kinematics,model based on gaussian process and model based on LSTM.Through ablation experiments on the I3AP model,the interaction module,maneuver recognition module,and attention module in the I3AP model can reduce the horizontal and vertical ADE errors to a certain extent,verifying the effectiveness of each module.Finally,in the scene verification based on the NGSIM-US101 data set,the I3AP model proposed in this paper has strong robustness and can accurately predict the future trajectory of various intentions.