VISUAL LOCALISATION OF UAV BASED ON DYNAMIC SEGMENTATION AND MULTI-LEVEL OPTIMISATION

Jianyu Chen, Wenhao Zhou, Wei Xiong, Jinwen Hu, and Zhao Xu

Keywords

UAV, autonomous localisation, dynamic feature point segmentation,visual SLAM

Abstract

The widespread adoption of unmanned aerial vehicles (UAVs) for inspection tasks highlights the demand for accurate autonomous localisation. Traditional visual simultaneous localisation and mapping (SLAM) methods often fail in dynamic environments due to interference from moving objects. On the one hand, learning-based methods are computationally expensive for onboard processing; on the other hand, purely geometric approaches lack adaptability. To address these issues, a visual localisation framework for inspection UAVs based on dynamic segmentation and multi-level optimisation is proposed. The method integrates optical flow and geometric constraints to eliminate dynamic features to ensure only static background features are used for localisation and mapping. In addition, the static feature extraction algorithm is enhanced through a score-based selection mechanism, which improves the accuracy of the front-end odometry and reduces cumulative error. Evaluations on public datasets and real UAV experiments show that the proposed approach significantly improves robustness and localisation precision, which is reduced by 17.80% on average compared to ORB-SLAM3 and DynaSLAM.

Important Links:

Go Back