机构地区:[1]上海理工大学管理学院,上海200093 [2]上海理工大学智慧城市交通研究院,上海200093 [3]上海理工大学智慧应急管理学院,上海200093 [4]上海理工大学光电信息与计算机工程学院,上海200093
出 处:《公路交通科技》2025年第3期1-10,共10页Journal of Highway and Transportation Research and Development
基 金:国家自然科学基金项目(52372304);教育部人文社会科学青年基金项目(22YJC790189);上海汽车工业科技发展基金会课题(2401);上海纽约大学上海市城市设计与城市科学重点实验室开放课题基金项目(2023 YWZhou_LOUD);上海理工大学智慧应急管理学院培育课题(24SIEMPY_ZYW)。
摘 要:【目标】为了提高复杂城市道路场景下智能网联汽车轨迹的长时预测精度,综合考虑了不同类型交通参与者(如车辆、行人、障碍物)及轨迹预测的不确定性。【方法】首先,提出了基于图遍历策略的考虑城市多类交通参与者的多模态轨迹预测模型,依据城市结构化道路场景和高清地图构建车道图,提取道路场景信息。其次,采用门控循环单元对目标车辆及其周边车辆历史轨迹和道路场景信息进行编码,并通过自注意力机制获取目标车辆与不同类型交通参与者及车道图信息间的交互特征。随后,通过图注意力网络处理,生成反映上下文信息的向量。然后,通过学习图离散遍历策略的不同可能性,结合轨迹预测模块在车道图的子集上探索多目标策略,输出多模态的预测结果。最后,基于凝聚层次聚类的轨迹聚类算法,将模型输出的1000条多模态预测轨迹进行聚类分析,选出输出概率最高的预测轨迹。【结果】所提出的轨迹预测模型在nuScenes公开数据集上的试验结果表明,其在评价指标最小平均位移误差和最小最终距离误差上相比其余主流的5种模型均有所降低,在6 s预测时域内展现了良好性能,使用RTX 4060在测试集上对每个样本进行预测的平均时间为0.017 s,进一步提升了模型性能。【结论】本研究方法适用于多类型交通参与者的复杂城市交通环境,有效提升了车辆轨迹预测精度和质量,且模型所需算力较低。[Objective]To improve the long-term prediction accuracy of intelligent connected vehicles’trajectories in complex urban road scenarios,this study takes into account the different types of traffic participants(e.g.,vehicles,pedestrians,obstacles)and the trajectory prediction uncertainty.[Method]First,considering multiple types of traffic participants,the multi-modal trajectory prediction model,based on the graph traversal strategy,was proposed.Based on the urban structured road scene and high-definition map,the lane graph was constructed,and the road scene information was extracted.Second,the gated recurrent units were used to encode the historical trajectories and road scene information of target vehicle and its surrounding vehicles.The interaction features among the target vehicle,different traffic participants,and lane graph information were obtained through the self-attention mechanism.Subsequently,processed with graph attention network,the vector reflecting context information was generated.Then,by learning the different possibilities of graph discrete traversal strategies and combining the trajectory prediction module,a multi-objective strategy was explored on a subset of lane graph,producing the multi-modal prediction result.Finally,more than 1000 multi-modal prediction trajectories output with the model were analyzed by using the trajectory clustering algorithm based on agglomerative hierarchical clustering.The predicted trajectories with the highest output probability were selected.[Result]The experimental results on public dataset nuScenes demonstrate that the proposed trajectory prediction model shows good performance within a 6-second prediction horizon,with lower minADE and minFDE scores compared with other 5 mainstream models.Additionally,the model requires low computational resources,with an average prediction time of 0.017 seconds per sample on the test set by using RTX 4060.The model performance is further improved.[Conclusion]The proposed model is suitable for complex urban traffic environments
关 键 词:智能交通 多模态轨迹预测 图遍历策略 多类型交通参与者 智能网联汽车
分 类 号:U495[交通运输工程—交通运输规划与管理]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...