行人
计算机科学
弹道
一致性(知识库)
人工智能
数据挖掘
工程类
运输工程
物理
天文
作者
Mingxu Wang,Xinhua Zeng,Hong Peng,Weilong Lin,Chengxin Pang
摘要
For the domains of path planning, environment perception, and control in autonomous driving, pedestrian trajectory prediction is highly significant. Given the uncertainty and complexity of potential movement directions in crowded scenarios, coupled with strong spatial interactions and temporal dependencies, accurately predicting pedestrian trajectories presents a challenging task. However, previous works have often overlooked the consideration of temporal consistency, which could introduce disturbances and interfere with the accuracy of prediction results. In our proposed approach, we construct a framework based on both temporal and spatial Transformers (referred to as TSTSC) to model pedestrian spatial graphs and capture temporal-spatio interactions. Specifically, we introduce a self-consistency constraint module based on self-supervised learning to ensure temporal consistency within scene intervals, reducing the impact of disturbances. This study tests using popular public datasets (ETH and UCY) and compares against existing methods. Experimental results demonstrate a significant improvement in predictive accuracy compared to baseline models, validating the soundness of our hypothesis.
科研通智能强力驱动
Strongly Powered by AbleSci AI