Human-like motion planning of autonomous vehicle based on probabilistic trajectory prediction

Motion planning for autonomous vehicles becomes more challenging when both driver comfort and collision risk are considered. To overcome this challenge, a human-like motion planning strategy based on the probabilistic prediction in a dynamic environment is proposed. In this study, it is mainly concerned with the following three aspects: the probabilistic prediction of states of the surrounding vehicles, decision making of the optimal path with a cost function and speed planning based on the driver’s target speed. Firstly, the path generation is realized based on a fifth-degree polynomial and the desired path is optimized by a cost function with four performance indices, i.e., safety, consistency, smoothness, and distance from local path to global path. Secondly, collision detection is analyzed regarding the host vehicle and surrounding vehicles aspects. For the host vehicle, a path-tracking error relating to vehicle speed and road curvature is taken into account. For the surrounding vehicles, the probabilistic trajectory prediction is made by using the structural Long-Short Term Memory (LSTM) network. Next, the collision probability is assessed using the Monte Carlo method and the optimal path is selected through the probability threshold depending on driver styles such as a conservative or aggressive driver. Moreover, the human-like speed planning for longitudinal motion is implemented considering driver target speed in vehicle following and vehicle cut-in situations. Finally, the proposed human-like motion planning algorithm has been validated via Hardware-in-the-loop (HIL) tests. The simulation results have shown the effectiveness of dynamic obstacle avoidance with global path-tracking and speed-tracking with driver comfort. Parameter sensitivity analysis for cost function and speed planner is then performed. The sensitivity analysis and the results also illustrate the influence degree of various parameters on the planned trajectory, which would be conducive to further improving the algorithm performance in the future. With an appropriate selection of the weight ratio between safety and comfort proposed in this work, it is found that the driver’s comfort acceptance will be improved compared with the traditional deterministic motion planning algorithm.