考虑路侧感知限制的交叉口冲突风险监测效能虚拟评价方法  

Method for Virtually Evaluating Roadside Sensing Units'Risk Monitoring Performance at Intersections Considering Perceptual Constraints

作　　者：郑玉冰 马羊 冯忠祥 丁恒 祝凯 ZHENG Yu-bing;MA Yang;FENG Zhong-xiang;DING Heng;ZHU Kai(School of Automotive and Transportation Engineering,Hefei University of Technology,Hefei 230009,Anhui,China;Anhui Lufeng Transportation Technology Co.Ltd.,Hefei 231131,Anhui,China)

机构地区：[1]合肥工业大学汽车与交通工程学院,安徽合肥230009 [2]安徽庐峰交通科技有限公司,安徽合肥231131

出　　处：《中国公路学报》2025年第3期48-64,共17页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52402402);安徽省重点研发计划项目(202304a05020050);安徽省自然科学基金项目(2308085QE188)。

摘　　要：路侧感知单元(Roadside Sensing Units,RSU)视场范围有限,且交通环境中动、静障碍物易影响RSU的感知精度。因此,即使多个RSU可协同工作,RSU可获取的轨迹数据与真实轨迹数据仍存在偏差且随RSU布设参数而变化,进而导致RSU冲突风险监测效能的差异。面向实地建造前RSU的优化配置,基于真实轨迹数据构建了交叉口RSU风险监测效能的虚拟评价方法框架。模拟RSU实时风险监测的工作机制,利用实测轨迹数据建立基于Agent的交叉口多模式冲突风险分析方法。融合静态场景模型与真实轨迹数据,进行动态三维交通场景重构。在RSU模型参数化表征的基础上,构建RSU协同感知轨迹的仿真方法。针对连续仿真获取感知轨迹的耗时性问题,以原始轨迹与交叉口概率占用图先验信息为输入,以协同感知仿真获取的轨迹为输出,基于深度神经网络(DNN)建立了考虑感知限制的轨迹生成方法。分别基于原始轨迹与可感知轨迹计算多类别冲突风险的TTC指标,利用风险交叉熵度量RSU的风险监测效能。利用CitySim公开数据集Intersection A与E的轨迹数据对方法框架进行测试。结果表明:交叉口RSU的冲突风险监测效能与RSU的布设参数密切相关;在测试集中,采用DNN生成RSU感知轨迹的RMSE均值为0.353;在相同的测试条件下,基于DNN方法与基于协同感知仿真方法计算的风险交叉熵值呈显著正相关,优于不考虑感知限制的简单感知模型;在16 GB内存、Intel®Core^(TM) i7-12700H@2.30 GHz、RTX-3060计算设备上通过协同感知仿真与DNN生成单条轨迹的平均时间分别在100 s与10^(-2) s数量级。相比于已有研究方法,所提方法框架考虑了RSU感知限制因素与交叉口多类别风险,可在RSU实地布设前量化评价风险监测效能,对辅助交叉口智能化改造具有积极意义。The field-of-view of a roadside sensing unit(RSU)is limited.In addition,static and dynamic occlusions in traffic scenarios may adversely affect the perception accuracy of an RSU.In such cases,even if multiple RSUs cooperate,the RSU-captured trajectories may deviate from the ground truth and vary with the RSU placement,thereby impairing the performance of the RSU in conflict risk monitoring.Considering the need to optimize the RSU configuration prior to real-world deployment,this study proposes a framework for virtually assessing the RSU risk monitoring performance at intersections using trajectory data.To emulate the real-time risk surveillance process of an RSU,an agent-based method was developed to identify and analyze multimodal conflict risks at intersections.The trajectory data were integrated with a three-dimensional intersection model to reconstruct dynamic traffic scenarios.After the parameterization of the RSU model,a virtual procedure was established to simulate cooperative trajectory perception.To address the time efficiency issue of collecting cooperatively perceived trajectory data using continuous simulations,a deep neural network(DNN)was introduced to generate co-perceived trajectories using prior knowledge(raw trajectory data and probabilistic occupancy maps)and simulated trajectory data as the input and response,respectively.The raw and RSU-captured trajectory data were used separately to compute the time-to-collision(TTC)indices of conflict events,and TTC-based cross-entropy was introduced to quantify the risk monitoring performance of the RSU.The proposed framework was tested on trajectory data collected from Intersections A and E in the open-source CitySim dataset.The results indicate that RSU risk monitoring performance is closely associated with RSU placement.The average root mean square error of the trained DNN model on the test dataset was 0.353,which implies that the DNN model can generate RSU-captured trajectories.The TTC-based cross-entropy obtained by the DNN-based method was significa

关 键 词：交通工程 风险监测效能 虚拟评价方法 路侧感知 交叉口 冲突风险 

分 类 号：U491[交通运输工程—交通运输规划与管理]