人机混驾环境下无信号交叉口自动驾驶汽车左转运动规划研究

为了使自动驾驶汽车在人机混驾环境下能安全、高效地左转通过无信号交叉口，在借鉴人类驾驶人左转时会对周围车辆驾驶意图进行提前预判的基础上，提出了一种基于周围车辆驾驶意图预测的自动驾驶汽车左转运动规划模型。首先将无信号交叉口处周围车辆的驾驶意图分为左转、右转、直行3种类型，利用相关向量机预测周围车辆驾驶意图，以概率形式输出意图预测结果并实时更新，进一步界定自动驾驶汽车与周围车辆的潜在冲突区域并判断是否存在时空冲突；接着，在充分考虑他车速度、航向及车辆到达冲突区域边界距离的基础上建立基于部分可观测马尔可夫决策过程的自动驾驶汽车左转运动规划模型，生成一系列期望加速度；最后，基于Prescan-Simulink联合仿真平台搭建无信号交叉口仿真场景，对所提左转运动规划方法进行仿真验证，将基于博弈论的运动规划方法、基于人工势场理论的运动规划方法与所提出的方法进行比较，并选取行进比例达到1所用的时间和碰撞次数作为评价指标。研究结果表明：基于相关向量机的驾驶意图预测方法可在自动驾驶汽车到达交叉口之前准确预测出他车驾驶意图；基于部分可观测马尔可夫决策过程的左转运动规划方法能够通过速度调整策略实现人机混驾环境下自动驾驶汽车与周围车辆在无信号交叉口处的交互；不同算法对比效果表明，所提左转运动规划方法在自动驾驶汽车与不同数量周围车辆交互的仿真场景下均可有效避免碰撞事故发生并提高自动驾驶汽车左转通过无信号交叉口的效率。

Left-turn Motion Planning of Autonomous Vehicles at Unsignalized Intersections in an Environment of Heterogeneous Traffic Flow Containing Autonomous and Human-driven Vehicles

To enable autonomous vehicles to turn left through unsignalized intersections safely and efficiently in an environment of heterogeneous traffic flow containing autonomous and human-driven vehicles, on the basis of learning the pre-judgment of the driving intention of surrounding vehicles when human drivers turn left, this paper proposes a left-turn motion planning model for autonomous vehicles based on the driving intention prediction of surrounding vehicles. First, the driving intention of surrounding vehicles at an unsignalized intersection was divided into three types: turning left, turning right, and going straight. A relevance vector machine was used to predict the driving intentions of the surrounding vehicles. The intent prediction result was produced in the form of probability and updated in real time. The potential conflict area between the autonomous vehicle and the surrounding vehicles was further defined, and whether temporal and spatial conflicts existed among the vehicles was determined. Next, considering the speed and heading angle of other vehicles and the distance between the vehicle and the boundary of the conflict area, a left-turn motion planning model for autonomous vehicles was established based on the partially observable Markov decision process (POMDP) to generate a series of expected accelerations. Finally, an unsignalized intersection simulation scene was built based on the Prescan-Simulink joint simulation platform. The proposed left-turn motion planning method was simulated and verified. Both game theory-based and artificial potential field theory-based motion planning methods were compared with the proposed method. The cost time for the travel ratio to reach 1 and the number of collisions were selected as evaluation indicators. The results show that the driving intention prediction method based on the relevance vector machine can accurately predict the driving intention of other vehicles before the autonomous vehicle reaches the intersection. The proposed left-turn motion planning method based on the POMDP can realize the interaction between an autonomous vehicle and surrounding vehicles at an unsignalized intersection through the speed adjustment strategy. The comparison results of various algorithms show that the proposed left-turn motion planning method can effectively avoid collisions and improve the efficiency of the autonomous vehicle's left-turning through unsignalized intersections in a simulation scenario where the autonomous vehicle interacts with a variable number of surrounding vehicles.