| 王海峰,桑海峰,王金玉,陈旺兴.结合速度控制的时空图网络行人轨迹预测模型[J].电子测量与仪器学报,2022,36(5):146-154 |
| 结合速度控制的时空图网络行人轨迹预测模型 |
| Spatial-temporal graph network with speed controlpedestrian trajectory prediction model |
| |
| DOI： |
| 中文关键词: 行人轨迹预测 生成对抗网络 速度控制 时空图网络 平均碰撞次数 |
| 英文关键词:pedestrian trajectory prediction generative adversarial network speed control spatial-temporal graph network average
collision times |
| 基金项目:国家自然科学基金（62173078）、辽宁省教育厅科研项目（LJGD2020006）资助 |
|
|
|
|
| 摘要点击次数: 1035 |
| 全文下载次数: 1242 |
| 中文摘要: |
| 行人轨迹预测中最重要的任务是建立行人轨迹交互模型,针对在模型中缺乏关于时间和速度等信息的建模,提出了一
种结合速度控制的时空图网络算法来建立行人交互模型并对轨迹进行预测. 整体模型采用条件生成对抗网络架构,其中采用速
度预测模块预测行人未来速度并作为条件生成对抗网络的控制条件,显式地将速度信息引入行人轨迹预测,避免较大偏差速度
对轨迹的影响。 在生成器中设计了基于图卷积注意力机制的时空信息融合模块,在提取行人轨迹序列运动特征并关注其空间
上相互作用关系的同时,显式地编码行人序列的时间相关性。 最后,将结合时空信息和速度信息的轨迹交互特征解码,完成轨
迹的预测。 此外,考虑到现有评价方法的不足,采用平均碰撞次数作为轨迹合理性的评判。 在公开数据集 ETH 和 UCY 上进行
验证,实验结果表明,该文所提出的算法能更好地完成行人轨迹预测,平均位移误差为 0. 40 m 和最终位移误差为 0. 79 m。 |
| 英文摘要: |
| The most important task in pedestrian trajectory prediction is to establish a pedestrian trajectory interaction model. Aiming at
the lack of semantic information about time and speed in the model, a spatial-temporal graph network algorithm combined with speed
control is proposed to establish pedestrian interaction model and predict trajectory. The overall model adopts the conditional generative
adversarial networks architecture, in which the speed prediction module is used to predict the future speed of pedestrians, and the control
condition of the conditional generative adversarial networks. The speed information is explicitly introduced into the pedestrian trajectory
prediction to avoid the influence of large deviation speed on the trajectory. A spatial-temporal information fusion module is designed in
the generator. While extracting the motion features of pedestrian trajectory sequence and paying attention to its spatial interaction, it
explicitly encodes the temporal correlation of pedestrian sequence. Finally, the trajectory interactive features combined with space-time
information and speed information are decoded to complete the trajectory prediction. In addition, considering the shortcomings of the
existing evaluation methods, the average collision times is used as the evaluation of trajectory rationality. The model is verified on the
public datasets ETH and UCY. The experimental results show that the proposed algorithm can better complete the pedestrian trajectory
prediction, with an average displacement error of 0. 40 m and a final displacement error of 0. 79 m. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
