UAV 추격 제어를 위한 미학습 타겟 대상의 마커리스 포즈 추정

초록

Unmanned aerial vehicles (UAVs) are effective for surveillance tasks involving tracking ground targets. Although artificial markers are widely used in various studies, they are unsuitable for most real-world applications. Moreover, most previous vision models are limited to seen targets which are already used for training. To address these challenges, this paper proposes a new markerless visual tracking and pose estimation framework for UAV tracking. In particular, Siamese-based tracking algorithm (NanoTrack) and video object segmentation model (XMem) were integrated for the visual tracking. A reliability-based feedback mechanism between these modules enables robust target recovery under severe occlusions, sudden background changes, and temporary frame losses. The visual tracking results were fused with other sensor signals (i.e., IMU, 1D-LiDAR, GPS) using an extended Kalman filter. Subsequently, the UAV was controlled to chase the moving target using a model predictive control. This framework was validated through real-world flight experiments, where the UAV successfully chased a ground vehicle navigating various trajectories such as straight and circular paths.

키워드