无保护左转场景下冲突车辆意图识别与一致性分析

Recognizing Intentions of Conflicting Vehicles and Analyzing of Intention Consistency in Unprotected Left-Turn Scenarios

摘要:

探究无保护左转场景下人类驾驶员意图的一致性方向和演化过程，以加快自动驾驶汽车在与人类驾驶汽车交互时双方意图一致性的达成。首先，基于中国上海仙霞剑河交叉口（XXJH）数据集和德国inD数据集，提取出无保护左转场景的实证轨迹。然后，基于同理心原则，提出合作加速度指标表征实时意图倾向，并利用支持向量机构建合作加速度、到达冲突点时间(T)和单边预期一致性方向的函数关系。接着，为探究双边预期一致性方向及其变化过程，将上述函数映射到左转车和对向直行车到达冲突点时间（即Tl和Ts）组成的坐标空间上，提出了决策时间域图的分析方法。在实际数据中挖掘人类驾驶员意图演化过程，得到预期意图冲突区域（即左转场景的“两难区”），在inD数据中两难区位于直线Ts=Tl上方，而在XXJH数据中预期冲突区分布整体位于此直线附近，从意图决策和演化的角度说明了inD数据中直行车具有更高的通行优先权，而XXJH数据中左转和对向直行车均认为双方通行优先权是相当的；挖掘了意图一致达成的置信区域，当交互状态位于在某一方到达冲突点时间小于2 s的区域，趋同比例平均为95%以上，可视为意图收敛区域。最后，讨论了将意图识别与一致性演化结果作为先验知识应用到自动驾驶汽车的交互策略，实现预期协同决策规划。

Abstract:

Unprotected turns account for efficiency loss and even accidents for autonomous vehicles at intersections. Exploring the consistency direction and evolution process of human driver intentions in this scenario can expedite the achievement of mutual intention consistency between autonomous and human-driven vehicles， thereby enhancing safety and efficiency. This paper aims to explore the orientation of intention consistency and evolution of human driver intentions in unprotected left-turns to accelerate the mutual intention consistency between autonomous and human-driven vehicles. First， based on the XXJH dataset from the Xianxiajian Intersection in Shanghai， China， and the inD dataset from Germany， empirical trajectories of unprotected left-turn scenarios were extracted. Then， based on the principle of empathy， an indicator named cooperative acceleration was proposed to represent real-time intent inclination. The relationship between cooperative acceleration， time to conflict point （T）， and unilaterally expected consistency direction was modeled using support vector machines. Furthermore， to explore the bilateral expected consistency orientation and its variation process， the aforementioned functions were mapped onto a coordinate space composed of the time to conflict point for left-turning vehicle and the opposite straight-going vehicles （i.e.， Tl and Ts）. An analysis method called time-domain-based decision diagram was proposed. By mining the evolution process of human driver intentions in empirical data， the conflicting region of expected intentions （referred to as the “dilemma zone” in left-turn scenarios） was identified. In the inD dataset， the dilemma zone was located above the line Ts=Tl, while in the XXJH dataset， the overall distribution of the expected conflict zone was near this line. It indicated from the perspective of intent decision and evolution that straight-going traffic had higher priority in the inD dataset， while in the XXJH dataset， both left-turn and opposing straight-going traffic considered their priority to be roughly equal. The confidence time domain for consistency were mined， i.e.， the interaction state was in an area where the time to conflicting point was less than 2s and the convergence ratio was on average above 95%. Finally， this paper treated the results of intention recognition and consistency evolution as a priori knowledge to the interaction strategy of autonomous vehicles and discussed the intended cooperative decision and planning.