A clustering approach for detecting moving objects captured by a moving aerial camera
This paper presents a novel method for detecting motion regions in image sequences obtained by rapidly moving cameras in fully 3-D scenes. The proposed method mainly focuses on the situations that the backgrounds of the image sequences change rapidly. It has three innovations over existing methods: First, it presents a new initialization method to fast and sparsely gather information of the background model, while traditional methods utilize a complicated training step. Second, a novel model updating scheme is proposed for establishing the on-line sparse background model iteratively. This is the main contribution of the proposed method and this enables the method to work in a 3-D scene which totally changes through the image sequence, while most other methods can only work with a pre-modeled scene or with a camera that moving in limited scope. Third, a novel two-stage model of the background and foreground motion regions is proposed and the foreground motion regions are detected using maximum a posterior estimation. The method is tested on various challenging image sequences captured by freely moving cameras and results show that it is very effective and robust.