Pedestrian trajectory prediction via physical-guided position association learning
Pedestrian trajectory prediction possesses huge application value in automatic driving, robots, and video surveillance. Due to the complexity of the environment and the uncertainty of pedestrians, predicting pedestrian trajectories is a challenging task. Previous studies simply employ the LSTM or tr...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-04-01
|
| Series: | Engineering Science and Technology, an International Journal |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2215098625000631 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Pedestrian trajectory prediction possesses huge application value in automatic driving, robots, and video surveillance. Due to the complexity of the environment and the uncertainty of pedestrians, predicting pedestrian trajectories is a challenging task. Previous studies simply employ the LSTM or transformer structure to construct the deep model, which hardly adequately mines the dependency relationship among different pedestrian positions from different views. In addition, directly employing the deep model to output the prediction results is easy to be disturbed by the external factor. To this end, we propose the Physical-guided Position Association Learning (PPAL) method to adequately explore the inter-position dependency relationship with the guidance of the physical motion rule. Specifically, to build the long/short-distance relationship, we develop the position association learning module (PAL) to deeply correlate different position coordinates by utilizing the advantages of the LSTM and transformer structure, which could stimulate the deep model to better perceive the pedestrian intention. In addition, the future motion trajectory has a strong correlation with the previous position and speed. Its physical motion rules provide much prior knowledge and increase the reasonability of trajectory predictions. Hence, we design the physical position modeling (PPM) to utilize the motion rule for trajectory prediction. Finally, we integrate PAL and PPM into a framework to deeply learn the inter-position dependency relationship. Abundant experiments on three mainstream databases demonstrate that the proposed PPAL significantly improves the prediction performance and surpasses other advanced methods. A large number of quantitative analyses show that the predicted trajectory is very close to the real trajectories, indicating that the proposed method has a better forecasting ability. |
|---|---|
| ISSN: | 2215-0986 |