Real-Time Temporally Consistent Monocular 6D UAV Pose Estimation for Onboard Aerial Perception
TL;DR
AeroMotion6D is proposed, a temporal transformer-based framework for monocular UAV 6D pose estimation from RGB video that consists of an adaptive context fusion mechanism that can incorporate past context information into the current estimation process and a persistent pose memory module that can convey pose-related information in two consecutive frames.
Abstract
The precision of 6D pose estimation is crucial for autonomous UAV perception, tracking, and navigation. Recent monocular pose estimation methods have shown encouraging results, but they are based on individual frames and do not fully utilize the temporal continuity in video sequences. As a result, even though pose estimation can be performed in a monocular manner, it can suffer from temporal jitter, unstable trajectories, and orientation ambiguity in fast motion, partial occlusions and challenging perspectives. To address these problems, AeroMotion6D is proposed, a temporal transformer-based framework for monocular UAV 6D pose estimation from RGB video. The suggested framework consists of an adaptive context fusion (ACF) mechanism that can incorporate past context information into the current estimation process and a persistent pose memory (PPM) module that can convey pose-related information in two consecutive frames. A symmetry-aware learning strategy is created to resolve orientation ambiguities in partially symmetric UAVs, and a motion-aware learning objective is created to promote pose evolution over time. AeroMotion6D continuously outperforms representative state-of-the-art techniques, according to experimental evaluations on public benchmarks; it achieves a mean absolute rotation error (MAEr) of 15.92∘ and a translation of 0.202 m on the DroneKey benchmark, and an average precision (AP) of 98.12% with a strict 10∘/10 cm success rate of 75.84% on the MAV6D benchmark. Furthermore, real-world validation on a physical Quanser QDrone platform confirms high robustness and practical applicability, yielding an average rotation error of 11.42∘, an average translation error of 0.141 m, and a 10∘/10 cm success rate of 87.53%. Embedded implementation experiments using an NVIDIA Jetson Orin NX with TensorRT FP16 optimization achieve real-time operation at approximately 17 FPS with an end-to-end latency of ∼58 ms per frame, demonstrating the practical onboard applicability of the proposed framework.