Fast and accurate visual tracking of ground buildings can provide unmanned aerial vehicles (UAVs) with rich perceptual information, which is very important for target recognition, navigation and system control. However, when an UAV moves fast, both background and buildings in visual scenes change relatively and rapidly. Consequently, there are no constant features for objects' appearance, which poses great challenges for visual tracking of buildings. In this paper, we first build an image manifold of buildings, which can encode the continuous variation of appearance. We then propose an efficient approach to learn this manifold and obtain more robust feature extraction results. By using a simple tracking framework, we successfully apply the extracted low-dimensional features to real-time building tracking. Experimental results demonstrate the effectiveness of the proposed method.
Online learning-based robust visual tracking for autonomous landing of Unmanned Aerial Vehicles