| 基于人-车交互的行人轨迹预测 |
| |
| 引用本文： | 连静,王欣然,李琳辉,周雅夫,周彬.基于人-车交互的行人轨迹预测[J].中国公路学报,2021,34(5):215-223. |
| |
| 作者姓名： | 连静 王欣然 李琳辉 周雅夫 周彬 |
| |
| 作者单位： | 1. 大连理工大学 汽车工程学院, 辽宁 大连 116024;2. 大连理工大学 工业装备结构分析国家重点实验室, 辽宁 大连 116024 |
| |
| 基金项目： | 国家自然科学基金项目（51775082，61976039）；中央高校基本科研业务费专项资金项目（DUT19LAB36，DUT20GJ207） |
| |
| 摘 要： | 针对行人轨迹预测具有复杂、拥挤的场景和社会交互问题，基于长短时记忆网络（Long Short-term Memory Network， LSTM）对行人与车辆、行人与其他行人的交互进行建模，提出一种基于人-车交互的行人轨迹预测模型（VP-LSTM）。该模型同时考虑了行人与行人的交互、行人与车辆的交互，更适用于复杂的交通场景。所构建的VP-LSTM包括3个输入，以行人的方向和速度作为历史轨迹序列输入，行人与行人的相对位置作为人-人交互信息输入，行人与车辆的相对位置作为人-车交互信息输入。该方法首先设计扇形人-人交互邻域和圆形人-车交互邻域来准确捕捉对被预测行人有相互作用的行人和车辆；其次建立3种不同的LSTM编码层来编码历史行人轨迹序列、人-人、人-车社交信息；然后定义人-人、人-车交互的防碰撞函数和方向注意力函数作为人-车、人-人社交信息的权重，进一步提高社会信息的精度；再将人-人、人-车交互信息输入到注意力模块中筛选出对行人影响大的社会信息；最后将筛选后的社会信息与行人历史轨迹序列一起输入到LSTM神经网络中进行行人轨迹预测，并在构建的DUT人-车交互数据集上验证提出的网络。研究结果表明：提出的方法能够准确地预测出交通场景中，人-车交互行人未来一段时间内的运动轨迹，有效提高了预测精度，提高了智能驾驶决策的准确性。
|
| 关 键 词： | 汽车工程 行人轨迹预测 长短时记忆神经网络 人-车交互 深度学习 注意力机制 |
| 收稿时间： | 2020-02-15 |
Pedestrian Trajectory Prediction Based on Human-vehicle Interaction |
| |
| LIAN Jing,WANG Xin-ran,LI Lin-hui,ZHOU Ya-fu,ZHOU Bin.Pedestrian Trajectory Prediction Based on Human-vehicle Interaction[J].China Journal of Highway and Transport,2021,34(5):215-223. |
| |
| Authors: | LIAN Jing WANG Xin-ran LI Lin-hui ZHOU Ya-fu ZHOU Bin |
| |
| Affiliation: | 1. School of Automotive Engineering, Dalian University of Technology, Dalian 116024, Liaoning, China;2. State Key Laboratory of Structure Analysis for Industrial Equipment, Dalian University of Technology, Dalian 116024, Liaoning, China |
| |
| Abstract: | A pedestrian trajectory prediction model (VP-LSTM) based on a long short-term memory (LSTM) network was proposed in this study to model the interaction between pedestrians and vehicles that can fit to complex and crowded scenes and social-interaction problems of pedestrian trajectory prediction. The model considers the interaction between pedestrians and vehicles, which is suitable for complex traffic scenarios. The VP-LSTM included three inputs: the direction and speed of pedestrians as historical track sequence inputs, the relative position of pedestrians as human-human interaction information input, and the relative position of pedestrians and vehicles as the human-vehicle interaction information input. First, the fan-shaped human-human and circular human-vehicle interaction neighborhoods were designed to accurately capture the pedestrians and vehicles that interacted with the predicted pedestrians. Second, three LSTM coding layers were established to encode the historical pedestrian-track sequence and the human-human and human-vehicle social information. Third, the anticollision function and direction attention function of human-vehicle and human-human interaction were defined as the weights of human-vehicle and human-human social information, respectively, to improve the accuracy of social information. Then, the information of human-human and human-vehicle interactions was inputted into the attention module to obtain the information focused on by that pedestrians. Finally, the filtered social information and pedestrian history track sequence were inputted into the LSTM neural network to predict the pedestrian trajectory. Finally, the DUT human-vehicle interaction dataset constructed by our group was used to verify the proposed network. The experimental results show that the proposed method can accurately predict the future movement-trajectory of pedestrians in traffic and improve the accuracy of intelligent vehicle decisions. |
| |
| Keywords: | automotive engineering pedestrian trajectory prediction LSTM human-vehicle interaction deep learning attention mechanism |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《中国公路学报》浏览原始摘要信息 |
|
点击此处可从《中国公路学报》下载全文 |
|
