Research On Planning And Motion Prediction For Intelligent Vehicles

Autonomous driving technology has the potential to reduce traffic accidents,improve the efficiency of transportation systems,and meet individual travel needs.Despite the great progress in autonomous driving technology in recent years,the safety and smoothness of autonomous driving still need to be improved in complex dynamic traffic scenarios.The planning system plays an important role in ensuring the safety and smoothness of autonomous driving,but the existing planning methods are not adaptable to complex dynamic environments.Predicting the motion of dynamic surrounding vehicles can improve the adaptability of the planning system,and ultimately improve the safety and smoothness of autonomous driving.This paper focuses on the planning technology of autonomous driving and motion prediction technology of dynamic surrounding vehicles.The main results and innovations are as follows:In order to understand the driving intention of surrounding vehicles in dynamic environments,an intention prediction method based on a probability framework is proposed.The method uses three types of time-serial driving characteristics to predict driving intentions,including physical states of vehicles,road structure and interaction behavior between vehicles.According to the causal relationship of features and the driving pattern of human drivers,an intent prediction network based on the dynamic Bayesian network is designed.In order to obtain states of features in the intent prediction network,a feature extraction algorithm is designed.Moreover,the selection algorithm of discretized parameters is also presented based on the influence strength.After obtaining states of all features in the network,the network parameters are learned,and the driving intention is probabilistically inferred.The prediction effect of the intent prediction method was verified by the real-traffic dataset NGSIM under the highway scene.The experimental results show that the F1 value can reach 0.795 and the predicted advance time can reach 3.75 seconds.The F1 value is improved by 30.33% compared with the traditional Bayesian network-based method.The predicted time is advanced by 2.72 seconds,2.25 seconds and 1.75 seconds compared with the traditional Bayesian network-based method,the hidden Markov model-based method,and the rule-based method,respectively.As for the low adaptability to the environment and low prediction accuracy of existing trajectory prediction methods,a trajectory prediction method based on an improved dynamic window method is proposed.The method uses an improved dynamic window method to predict the trajectory of the vehicle based on the vehicle kinematics.In order to improve the accuracy of the trajectory prediction,the predicted intention is used as a priori knowledge to reduce the size of the dynamic window.In order to consider the interaction of environmental vehicles,the cost functions of cooperative interaction and interfering interaction are designed separately.In addition to training and testing on the real-traffic dataset NGSIM,the method is also tested in dynamic scenes built by the simulated platform Prescan.The experimental results show that the proposed trajectory prediction method has the capacity to work in scenes without training samples,thus improving the adaptability to the environment.The root mean square error and mean absolute error of the predicted lane-change trajectory on NGSIM are reduced by 8.98% and11.20% compared with the constant velocity based method.Under the simulated experimental environment,the root mean square error of predicted trajectory within 2-3 seconds is reduced to 63.78%-67.71% of constant velocity based method.In order to improve the safety and smoothness of the autonomous driving in uncertain environments,a planning method based on the trajectory prediction of other dynamic vehicles is proposed.The planning method decouples the planning system into path planning and speed planning,and considers the motion impact of dynamic vehicles.The path planning adopts the geometric symmetry configuration method to generate a path set,which conforms to the kinematic constraints of the autonomous vehicle and its state space.The method improves the real-time performance of generating paths.The performanceevaluation function is designed to obtain the final path based on the collision probability of the ego path set and the predicted trajectory of dynamic vehicles.In order to avoid the speed planning being too conservative,the collision intersection and the safety distance are determined to planning the desired speed of the ego vehicle.At last,the performance of the proposed planning method is verified in Prescan.The experimental results show that the proposed method drives toward a safer area under the experimental scene.Moreover,curvature changes of the path and the speed profiles are smoother than the traditional planning method without trajectory prediction.