Trajectory prediction is critical for autonomous systems such as self-driving cars and surveillance systems. In this work, the objective is to predict the future paths of road users by analyzing their past movement patterns and the surrounding environment. Using video data, we investigate the influence of perspective on the effectiveness of human trajectory prediction in two key scenarios: knowledge transfer from a source dataset to a target dataset and training models exclusively on the target dataset. Findings reveal that aligning dataset perspectives is critical for optimal transfer learning performance. In particular, prediction accuracy improves with increasing pitch angles, highlighting the importance of this parameter in trajectory modeling. A drone-type view, offering a bird’s-eye perspective, further enhances prediction quality by capturing spatial relationships more comprehensively. Additionally, the inclusion of the estimated homography significantly improves the prediction performance by refining the representation of spatial data.
Martins et al. (Fri,) studied this question.