This paper presents an autonomous computer vision system for tracking multiple dynamic ground targets, from images acquired by a camera onboard of a μ-UAV. The method proposed is a self adaptive technique that seamlessly integrates ego-motion compensation with target detection and tracking to provide robust localisation of ground targets. Ego-motion compensation is achieved through establishing homographies using target independent invariant feature descriptors. Targets are then detected using a novel background learning strategy where the optical flow field is fused together with a dynamic background model for accurate foreground extraction. In addition, the paper also reports the use of a Monte Carlo joint probabilistic data association filter for tracking multiple unknown targets. The field tests demonstrate the capabilities of the vision system based on experimental results on images captured by a camera on-board of quadrators (μ-UAV).
展开▼
