车辆-行人运动不确定条件下虚警率约束的行人防撞预警模型研究

Pedestrians are vulnerable road use groups in modern traffic, and pedestrian perception and its anti-collision warning are one of the key common technologies to be solved for the development of intelligent vehicles. The project is innovatively proposed, which can identify the intention of pedestrian movement, and build an adaptive decision-making model for pedestrian collision warning system with controllable false alarm rate. Based on deep convolution network, pedestrians and their skeletons can be detected and located, and take the pose vector of key skeleton feature point in single frame as a link, A spatio-temporal correlation model of pedestrian attitude and movement intention can be build based on recurrent neural network, which can reveal path pattern of pedestrian movement. Through the fusion method of confidence data in pedestrian detection, tracking, location and intention identification, the probability distribution model of pedestrian path cluster can be established. Based on vehicle system dynamics, the joint probability distribution of vehicle motion reachable domain is studied, then a comprehensive probability distribution model of vehicle-pedestrian collision is built, which can provide conditions for the modeling of index variables of subsequent warning decision-making. Based on driver's actual handling data, A parameterized model of accurate warning decision index can be established. An adaptive decision-making triggering model for pedestrian collision warning can be conducted with constraints on false alarm rate, and followed by verification of hardware-in-loop and real-car test. The research results can provide theoretical basis for the deep environment perception and adaptive decision making technology of intelligent vehicles, which is of important scientific and engineering significance.

行人是现代交通的易受伤害道路使用群体，行人感知与防碰撞预警是智能车辆发展亟待解决的重大关键共性技术之一。本项目提出一种辨识行人运动意图且虚警率可控的行人防碰撞预警自适应决策模型构建的新方法：基于深度卷积网络检测定位行人及其骨架特征点，以帧内关键骨架特征点姿态向量为纽带，基于递归神经网络建立行人姿态与运动意图的时空关联模型，揭示行人运动路径模式；采用行人检测跟踪定位与意图辨识中的置信度数据融合方法，建立行人运动路径簇的概率分布模型，并基于车辆系统动力学，研究车辆运动可达域的联合概率分布，最终获得人车空间碰撞的综合概率分布模型，为后续预警决策指标变量建模提供条件；基于驾驶员实操数据，建立准确的预警决策指标的参数化模型，最终构建虚警率约束的行人防撞预警自适应决策触发模型，并进行在环/实车实验验证。研究成果可为智能车辆深度环境感知与自适应决策技术提供理论依据，具有重要的科学和工程意义。

行人是交通流的易受伤害道路使用群体，行人感知与防碰撞预警系统的设计与研发一直是自动驾驶系统商业化的关键功能之一。本项目提出了一种辨识行人运动意图且虚警率可控的行人防碰撞预警自适应决策模型构建的新方法，该方法本质上属于数据驱动：基于深度卷积网络检测定位行人及其骨架特征点，以帧内关键骨架特征点姿态向量为纽带，基于递归神经网络建立行人姿态与运动意图的时空关联模型，揭示了行人运动路径模式与模态等；采用行人检测跟踪定位与意图辨识中的置信度数据融合方法，建立了行人运动路径簇的概率分布模型，并基于车辆系统动力学，研究了车辆运动可达域的联合概率分布，最终获得了人车空间碰撞的综合概率分布模型，为后续预警决策指标变量建模提供条件；基于上海松江采集的驾驶员实操数据，建立了准确的预警决策指标的参数化模型，最终构建了虚警率约束的行人防撞预警自适应决策触发模型，并进行在环平台的实验验证。研究成果表明，该算法具有较大的接受度且准确预警时机好，可为智能车辆深度环境感知与自适应决策技术提供理论依据，具有重要的科学和工程意义。取得的重要研究成果有：.1）建立了基于真实交通事故和CARLA软件的场景数据库，并结合社会力模型实现了行人在道路上的骨架关键点的多目标在线检测与运动跟踪，其MOTP值达到81.2%。.2）提出了一种用于多智能体轨迹预测的通用神经网络架构，可在车辆直行、左转及右转的情况下，实现车辆/行人等交通代理轨迹的高精度轨迹预测，且在车速为5m/s-20m/s等情况下实现了非预期行人碰撞风险评估。.3）综合考虑行人安全、车辆制动能力以及人机交互等机制，形成了行人防撞的分级预警策略机制，在30km/h的典型事故场景复现中可显著降低误触发率。