Research On Motion Analysis Of Traffic Objects For Autonomous Lane Changing

The lane-changing behavior of a vehicle driving in complex traffic environment is not only affected by internal factors such as the driver’s driving level and habits,but also by external factors such as the state of surrounding traffic vehicles,traffic flow density,and traffic signals.In the closed-loop interactive system of "human-vehicle-environment",the driver is the weakest link.Therefore,the goal of research on autonomous lane changing technology is to use the sensing system to enhance the driver’s perception and decisionmaking ability,and to improve the comfort,economy and safety of vehicle lane changing.At present,for the analysis and research of traffic vehicles in lane changing scenarios,more attention is paid to obtaining lane occupancy,relative distance and relative speed between the target vehicle and the vehicle based on sensor information as the basis for judging the timing of lane changing.However,under the working conditions such as curves and lane changes,it is impossible to accurately identify the working conditions of the target vehicle only with sensor information.At the same time,the related research pays less attention to the movement of the target vehicle in the lane,and accurate movement information is the premise to ensure the accuracy of traffic vehicle behavior recognition and trajectory prediction.Therefore,in order to improve the perception ability of the intelligent driving system under multiple working conditions,this paper conducts research on the traffic vehicle motion analysis method with robustness under various working conditions such as straight roads and curves.The research contents mainly include:First,research on multi-sensor target tracking and information fusion.In order to suppress sensor noise and realize circumferential perception,based on the multi-sensor arrangement scheme,this paper proposes a method of applying Kalman filtering to a single sensor for tracking,and using the covariance matrix obtained by tracking for multi-sensor fusion.The single sensor obtains the optimal estimation through the predicted value and sensor observation value of Kalman filter,which effectively improves the accuracy of sensor ranging.The global nearest neighbor algorithm is used for target matching among multiple sensors,and the covariance is interactively fused to realize the information perception and state management of circumferential traffic vehicles.Second,the traffic target vehicle movement analysis.On the premise of not adding sensors,this paper introduces a transformation model based on lane line information,and generates the backward lane line in combination with the ego vehicle motion,which can ensure the error and stability of the generated lane line.Aiming at the ambiguity introduced by the sensing information under curved road conditions,the horizontal and vertical trajectories decoupled from the ego-vehicle motion are obtained based on the historical information of the target vehicle,and the in-lane motion of the target vehicle is decomposed into vertical lane lines(horizontal)and along lane lines(longitudinal)in two dimensions.Compared with the analysis in the sensor coordinate system,the target object analysis in the center reference line coordinate system of the vehicle lane proposed in this paper is more efficient and accurate.Finally,the accuracy verification of the algorithm under multiple working conditions is completed in the simulation scenario.Thirdly,research on target vehicle behavior recognition and trajectory prediction.Using the acquired motion information in the target vehicle lane as the model input,the vehicle behavior recognition model based on continuous hidden Markov and the trajectory prediction model based on LSTM network are established respectively.The vehicle behavior recognition model can accurately identify each stage of the lane changing process through the lateral features in the motion information.At the same time,taking the trajectory information and motion information of the target vehicle as input,the trajectory prediction model based on LSTM network has better performance in long-term prediction than the kinematic model.Fourth,real vehicle data verification.A real vehicle test platform is built to verify the multi-sensor target tracking and fusion algorithm and the target vehicle motion analysis algorithm.Aiming at the research on the motion analysis of traffic vehicles under various working conditions,this paper determines a target vehicle motion analysis method with high robustness and applicability.Safety decision-making provides some support,thereby improving the safety and comfort of the driver’s lane-changing behavior.