一种基于矩阵记忆长短期记忆网络的行人轨迹预测算法OA
Pedestrian trajectory prediction algorithm based on matrix memory long short-term memory network
针对行人轨迹预测中存在的时序特征建模不足、多尺度融合缺乏明确区分以及多任务训练不稳定等问题,提出一种基于矩阵记忆长短期记忆网络(matrix long short-term memory,mLSTM)的纯时序预测算法.该算法构建以mLSTM为核心的编码器-解码器架构,挖掘轨迹的时间依赖特征;设计多尺度轨迹特征融合模块,采用双向策略实现短期与长期特征的层次化表达;引入指数移动平均标准化的多任务机制,提升训练的稳定性与模型的泛化能力.在ETH和UCY数据集上的实验结果表明,该算法相较于Trajectory-Transformer和SGCN,在平均位移误差上分别降低14.81％和16.21％,在最终位移误差上分别降低19.66％和4.62％,展现出良好的预测精度与鲁棒性,为行人轨迹预测提供稳健有效的基础模型.
This paper proposed a purely temporal prediction algorithm based on mLSTM to address insufficient temporal mode-ling,ambiguous multi-scale feature fusion,and unstable multi-task training in pedestrian trajectory prediction.The algorithm established an encoder-decoder architecture centered on mLSTM to capture temporal dependencies in trajectories.It designed a multi-scale trajectory feature fusion module with a bidirectional strategy to enable hierarchical representation of short-term and long-term features.This paper introduced an exponential moving average normalization-based multi-task mechanism to improve training stability and model generalization.Experimental results on the ETH and UCY datasets show that the proposed algo-rithm reduces average displacement error by 14.81％and 16.21％,and final displacement error by 19.66％and 4.62％,re-spectively,compared with those of Trajectory-Transformer and SGCN.The results demonstrate high prediction accuracy and robustness,providing a stable and effective backbone model for pedestrian trajectory prediction.
厍向阳;王艺龙
西安科技大学人工智能与计算机学院,西安 710600西安科技大学人工智能与计算机学院,西安 710600
信息技术与安全科学
行人轨迹预测矩阵记忆的长短期记忆网络多尺度特征融合指数移动平均多任务学习
pedestrian trajectory predictionmatrix long short-term memory networkmulti-scale feature fusionexponential moving averagemulti-task learning
《计算机应用研究》 2026 (2)
472-478,7
陕西省自然科学基础研究资助项目(2019JLM-11)陕西省科技计划资助项目(2021JQ-576)陕西省教育厅项目(19JK0526)
评论