交叉口环境下基于三维点云的目标车辆行为辨识与预警算法研究

The interweaving conflict of vehicles at different directions is one of the main causes of the frequent traffic accidents at intersections. Nowadays, autonomous vehicle mainly obtains the status of moving vehicles at intersections through on-board communication technology. But the popularization of vehicle networking technology has not yet been achieved due to technical standards and business models. In order to improve the redundancy of intersection environment sensing, this project will use 3D lidar to actively perceive the behaviors of multiple moving vehicles at intersections and execute the intelligent pre-warning strategy, independent of on-board communication equipment and roadside facilities. First, the point cloud density of long-range vehicle is increased on the basis of keeping the overall shape of vehicle point cloud unchanged, and the pose model of optimal rectangular is proposed by multi-frame fitting method. Second, moving vehicles with strong correlation constraints are screened using the intention of autonomous vehicles, and an enhanced tracking model with multiple motion states is established by using the state assumption of occluded vehicles and the constraint of variable elliptical tracking gate. Next, the motion characteristics of multiple driving behaviors are explored, and the online identification model of target vehicle behavior is built using machine learning algorithm. Finally, multi-level warning strategy of conflict risk is proposed through analyzing the whole driving situation of multiple vehicles. The research result will provide a new theoretical basis and technical support for environmental awareness system and facilitate the practical application of autonomous vehicle in complex road scenarios.

各方向车辆交织冲突是导致交叉口处交通事故频发的最主要原因之一。目前无人车主要通过车载通信技术获取交叉口运动车辆状态信息，但是受限于技术标准和商业模式等问题，车联网技术产业化普及尚未实现。本项目以提高交叉口环境感知冗余度为目标，拟在不依赖车载通信设备、路侧通信设施的情况下利用车载三维激光雷达对交叉口运动多目标车辆行为进行全方位主动辨识和智能化预警。在保持车辆点云整体形状不变的基础上增强远距离车辆点云密度，建立多帧拟合最优矩形位姿求解模型；筛选与无人车行驶意图呈强相关约束的多目标车辆，利用被遮挡车辆状态假设和可变椭圆跟踪门约束建立考虑多种运动状态的增强式跟踪模型；挖掘目标车辆不同行为工况的运动特性，运用机器学习算法建立目标车辆行为在线辨识模型，并协同耦合多车行驶态势提出冲突风险多级预警策略。研究成果将为环境感知系统的研发提供新的理论基础和技术支撑，有助于推动无人车在复杂道路场景下的实际应用。

车辆交织冲突是导致交叉口处交通事故频发的主要原因之一。本项目以提高交叉口环境感知可靠性为目标，在不依赖车载、路侧通信设备的情况下利用车载三维激光雷达对交叉口运动多目标车辆行为进行主动辨识和智能预警。首先在保持车辆点云整体形状不变的基础上增强点云密度，建立多帧拟合最优矩形位姿求解模型；利用被遮挡车辆状态假设和可变跟踪门约束建立考虑多种运动状态的跟踪模型；分析目标车辆不同行为特性并建立在线辨识模型，接着协同耦合多车行驶状态提出冲突风险预警策略。发表国际期刊和会议论文6篇，授权发明专利2项，出版专著1部，获省部级科技奖励1项，毕业硕士生3人。相关研究成果将为环境感知系统的研发提供新的技术支撑，有助于推动无人车在复杂道路场景下的实际应用。