Abstract.A model based predictive motion planning with combining Bayesian Methodology for autonomous driving is proposed. The planning algorithm firstly generates candidate obstacle free trajectories among the current state and sampled desired states using trajectory generation techniques. Then Bayesian Methodology is applied to select the optimal trajectory among candidate trajectories. The preferred trajectory selected in the previous planning cycle is regarded as the prior knowledge, while the trajectory cost in current cycle is transformed into likelihood function. Both the distribution follows the normal distribution and the Bayesian theorem is adopted to calculate the posterior knowledge to determine the preferred trajectory. Then the speed profile is calculated with the preferred trajectory to produce the real trajectory. The proposed motion planner is implemented and tested in simulations. Experiments results show that the planner has good performance in autonomous driving and especially reduces indecision behavior in uncertain environments, and improves stability of autonomous driving.
IntroductionThe motion planning is a fundamental technology in autonomous driving. The high performance of motion planner will improve greatly the performance of autonomous driving, especially stability, safety, smoothness, and speed. Because of the inherent dynamic-mechanism delay and hysteresis in autonomous systems, predictive approach [1, 2] is commonly utilized in motion planning. In this approach, motion planning has two key procedures: 1) trajectory generation, and 2) trajectory selection. The former generates motion planning search space, while the latter determines or selects the optimal or preferred trajectory.State space sampling method [3,4] and Frenet Frame based semi-reactive technique [5] is used to generate feasible trajectories on-line. They are both based on predictive control approach. Commonly, the performance of motion planner mainly depends on trajectory selection; however how to select the preferred trajectory is not mentioned clearly in above researches. In [5] the optimal trajectory was determined by utilizing optimal control approach, but the cost function was different in different behaviors which might make parameters tuning to be complex and superfluous.In this paper, an efficient motion planner combining trajectory generation using quintic polynomials with trajectory selection adopting Bayesian method is proposed. The planner has the following features: generated trajectories are satisfying differential, dynamic and road constraints, therefore are kinematically feasible, and are smooth and continuous enough due to high order polynomials; Bayesian based trajectory evaluation and selection deals with the indecision and unstable problem in planning, and guarantees time consistency with sequence knowledge which is formulated as probabilistic principles.