Authors:

Aakarsh Nagendra Rai

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Trajectory prediction in crowded environments has traditionally relied on single-view or overlapping multi-view setups. In this work, we explore a more challenging and realistic setting: forecasting the future trajectories of multiple agents observed by spatially distributed drones with non- overlapping fields of view (FoVs). Each drone independently observes a subset of the scene through a circular FoV and encodes its local spatio- temporal observations into a compressed representation using lightweight CNN encoders. These representations are transmitted to a central fusion model that spatially aligns and aggregates the embeddings to construct a unified global context. A Transformer-based decoder then jointly predicts future heatmaps of agent positions, reasoning over inter- agent dynamics across disjoint views. Our method effectively addresses issues of spatial disjointedness, partial observability, and communication constraints. We evaluate on a synthetic drone surveillance dataset and demonstrate strong performance in predicting trajectories even when agent paths traverse across different FoVs, highlighting the model's ability to learn implicit spatial coordination without explicit overlap or shared views Keywords: Machine Learning; Trajectory Prediction; Non-Overlapping Field of Views; Federated Learning

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Aakarsh Nagendra Rai