Improve the LSTM Trajectory Prediction Accuracy through an Attention Mechanism | |
Zhang, Tong; Wang, Zhiwen | |
2022 | |
会议日期 | JUN 15-17, 2022 |
会议地点 | Anaheim, CA |
DOI | 10.1109/ITEC53557.2022.9813863 |
页码 | 190-195 |
英文摘要 | Although the long-short term memory (LSTM) network has been widely adopted to predict the vehicle trajectory, the iterative nature of LSTM introduces the accumulative errors. The accumulative errors result in a gradual decrease in the accuracy of trajectory prediction over time. Therefore, how to reduce the accumulative errors of the LSTM is a very critical issue. To solve this problem, we introduce a two-stage attention mechanism with the LSTM Encoder-Decoder model, which uses the spatial attention mechanism and the output attention mechanism to weight the input hidden layer features and the output prediction of the decoder. In this way, the accuracy of trajectory prediction is improved. The cumulative error of the predicted trajectory is significantly reduced. The proposed method is validated on US-101 and I-80 datasets from NGSIM. The simulation results show that the test dataset's average error at one second and five seconds is reduced from 0.7676 meters and 7.7168 meters to 0.4601 meters and 4.2184 meters, respectively. The average prediction error is reduced by 45.33%. |
源文献作者 | IEEE,AIAA,IEEE Power Elect Soc,IEEE Ind Applicat Soc,IEEE Power & Energy Soc,IEEE Transportat Electrificat Commun |
会议录 | 2022 IEEE/AIAA TRANSPORTATION ELECTRIFICATION CONFERENCE AND ELECTRIC AIRCRAFT TECHNOLOGIES SYMPOSIUM (ITEC+EATS 2022) |
会议录出版者 | IEEE |
语种 | 英语 |
WOS研究方向 | Engineering ; Transportation |
WOS记录号 | WOS:000848063600033 |
内容类型 | 会议论文 |
源URL | [http://ir.lut.edu.cn/handle/2XXMBERH/159871] |
专题 | 电气工程与信息工程学院 |
作者单位 | Lanzhou Univ Technol, Lanzhou, Gansu, Peoples R China |
推荐引用方式 GB/T 7714 | Zhang, Tong,Wang, Zhiwen. Improve the LSTM Trajectory Prediction Accuracy through an Attention Mechanism[C]. 见:. Anaheim, CA. JUN 15-17, 2022. |
个性服务 |
查看访问统计 |
相关权益政策 |
暂无数据 |
收藏/分享 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论