轨迹拓扑引导的多智能体轨迹预测模型OA
Multi-agent trajectory prediction guided by trajectory topology
为解决现有轨迹预测方法难以准确理解交互关系与复杂场景信息的问题,提出轨迹拓扑引导的多智能体轨迹预测模型(trajectory topology-guided,TTG).采用时空特征编码方法对场景信息进行编码,利用拓扑融合器与拓扑解码器对编码特征进行提取,并计算交互概率.利用特征融合模块对未来预测轨迹进行场景道路约束,在Argovese运动预测基准数据集进行实验分析.结果表明:TTG的概率加权最小最终位移误差b-mFDE6为1.73,最小平均位移误差mADE6为0.76,相比基准模型SIMPL分别减少4.42％与3.8％.
Existing trajectory prediction methods usually struggle to accurately comprehend interactions and complex scene information.To address the issue,this paper proposes a multi-agent trajectory prediction model TTG(trajectory topology-guided).A spatiotemporal feature encoding method is employed to encode scene information.A topology fusion module and a topology decoder are employed to extract encoded features and compute interaction probability.A feature fusion module imposes scene road constraints on the future predicted trajectories.Experiments conducted on the Argoverse motion forecasting benchmark demonstrate TTGNet achieves a probability-weighted minimum final displacement error(b-mFDE6)of 1.73,down by 4.42％and a minimum average displacement Error(mADE6)of 0.76,down by 3.8％compared to those of the baseline model SIMPL.
邓召学;王金权;王戡;李兴泉
重庆交通大学机电与车辆工程学院,重庆 400074重庆交通大学机电与车辆工程学院,重庆 400074招商局检测车辆技术研究院有限公司,重庆 401122重庆长安汽车股份有限公司汽车工程研究总院,重庆 401120||重庆理工大学车辆工程学院,重庆 400054
信息技术与安全科学
轨迹预测多智能体特征编码轨迹拓扑
trajectory predictionmulti-agentfeature encodingtrajectory topology
《重庆理工大学学报》 2026 (9)
36-42,7
国家自然科学基金项目(52072054)
评论