Skip to content

Topologically Stable UAV Object Detection: Formal Robustness Guarantees via Persistent Homology #621

@kamel-houari

Description

@kamel-houari

AV object detection faces fundamental challenges from dense grid representations
that waste computation and provide no robustness guarantees. We propose NTEFNet,
where detection operates over topological graphs derived via persistent homology
(H0 ⊕ H1), enabling the �rst formal stability bound for UAV detection: degradation
under perturbation ε is bounded by O(ε). A Hierarchical Graph Pyramid reduces
4K latency from 312ms to 84ms on Jetson Orin AGX. On VisDrone-VID, NTEF-Net
achieves +8.1% recall in low-light (p < 0.01), +9.3% MOTA under occlusion, and
12.4% robustness advantage over grid-based methods across 75 corruption conditions,
while reducing FLOPs by 35%. Rigorous ablation con�rms gains stem from topological
invariance, not graph structure alone.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions