Research On Trajectory Planning Considering Motion Uncertainty Of Traffic Participants

The rapid development of the automobile industry has contributed to the economic growth of the country and provided convenient travel services to the population.At the same time,congested driving environments and various human factors have contributed to the high number of traffic accidents in the world over the years.How to assist human driving operations to improve active safety,or directly free humans from driving tasks,is what researchers expect from autonomous driving.As one of the key technologies of autonomous driving,trajectory planning is responsible for the safety risk assessment of the current scenario and planning a safe,comfortable driving trajectory.The intricacies of real scenario traffic,especially the strong uncertainty of motion about traffic participants,make decision planning safety for autonomous driving challenging.In order to improve the quality and safety of trajectory planning under obstacle avoidance scenario,the planning algorithm’s cognition of environment cannot stay at the level of obstacle detection and tracking,but needs to predict the motion of surrounding traffic participants,fuse the perception and prediction information to assess the scene safety risk posture,and plan the trajectory on this basis.With this research theme,this paper firstly establishes a prediction model based on deep learning to predict the motion trajectory of surrounding vehicles,and then establishes a trajectory planning module based on multi-objective optimization to realize trajectory planning in obstacle avoidance scenarios by combining perception and prediction information.Firstly,a prediction model based on deep learning is built.In response to the complexity and diversity of influencing factors,the scene is abstracted into a vectorized model to achieve a unified expression of information about the characteristics of environmental elements,which include static map elements and dynamic traffic participants.In order to make full use of the input information,an encoder with a hierarchical graph network structure is built to encode the context vector,first capturing local features of environmental elements by a multilayer perceptron,such as geometric features of map elements or motion features of traffic participants,and then capturing the interaction features between target agent and environmental elements by a multiheaded self-attention.For the uncertainty and multimodality of traffic actor motion prediction,the decoder uses a target-driven prediction framework.The sampling strategy is designed for target agent,and the candidate target points are sampled in the map to inject map information for the decoder,so that the decoder can decode the context vectors into multimodal predicted trajectories based on different targets and improve the multimodal prediction capability and accuracy of the model.The target point-driven prediction framework is divided into three phases to achieve trajectory prediction step by step.Firstly,the probability distribution and offset of the candidate targets are estimated to get a set of prediction modes of the target agent,then the trajectory sequence is estimated guided by the given prediction target points,and finally the prediction trajectories are scored and selected to output a set of prediction trajectories.Secondly,a path-velocity decoupling planning module based on multi-objective optimization is established.For the characteristics of direct trajectory planning with high dimensionality and spatio-temporal coupling,the technical process of path-velocity decoupled planning is proposed to reduce the complexity of the problem.In order to make the planning more natural and direct,a curve coordinate system is established as the baseline for planning.In the path planning step,path point sampling is performed to establish the topological structure graph for the non-convex characteristics of the solution space,and the dynamic planning algorithm is designed to search for the optimal path curve in the topological graph,which realizes the obstacle decision in the lateral direction and serves as a basis to open the sub-convex space containing the optimal decision in the non-convex space,creating the conditions for the subsequent path optimization.Then,a quadratic programming model is constructed on the convex space and solved to produce a path profile that satisfies the constraints and is multi-objective optimal.In the speed planning step,considering the uncertainty of traffic participants’ motion,the S-T diagram is established to predict the safety risk by integrating the prediction information,and the dynamic planning algorithm is designed to search for the optimal speed profile in the S-T diagram,which realizes the longitudinal obstacle decision and opens the optimal convex space,and then the quadratic planning model is constructed to optimize the speed,and the speed profile satisfying the constraint and multi-objective optimal is solved.The iterative optimization on path and velocity yields multi-frame smooth and stable planning trajectories.Finally,a simulation environment based on the dataset l5 kit was established with the scenario as the unit,and the training and testing of the prediction model were realized by extracting samples to make the training and testing sets,while the prediction performance was visualized and analyzed by conducting simulation experiments in several scenarios to verify the validity of the model.Then the simulation experiments of the planning module in the obstacle avoidance scenario are carried out by combining the perception and prediction information,and the multi-frame planning trajectory is smooth and stable,which verifies the effectiveness of the model.