DRL: An efficient heterogeneous spatial feature interaction framework for UAV self-localization

doi:10.1016/j.patcog.2026.113330

DRL: An efficient heterogeneous spatial feature interaction framework for UAV self-localization

2026-02-17

0

PRE

AI

DOI：10.1016/j.patcog.2026.113330

Original

Original request for help

Save

Abstract

En 中文

• We propose DRL, an efficient end-to-end UAV self-localization method that enables learnable heterogeneous feature interactions. • We design two DRL variants–Post-Fusion and Mix-Fusion–and analyze their respective advantages and limitations. • We establish a new benchmark with the UL14 dataset of paired samples and introduce two metrics: MA@K and RDS. • DRL achieves a 9.4% improvement in MA@20, while reducing inference time by over 7 ×  and storage by 3 × , reaching 100 FPS.

Keywords:

UAV self-localization

heterogeneous feature interaction

deep reinforcement learning

efficient framework

spatial feature processing

DRL: An efficient heterogeneous spatial feature interaction framework for UAV self-localization

Abstract

Journal

Organization