Detection of Objects in a scene is quite important nowadays, e.g. for surveillance purposes. Most of the algorithms used require a static background to be able to separate moving or added objects from the monitored area. If the camera is not static, knowledge about the camera movement is required to be able to compensate the global motion. If the watched scene can be approximated as planar, the compensation is quite easy using a affine or projective transformation. However, if the scene is not planar but contains 3D objects, like furniture, walls, or buildings, for instance, a more complex compensation is required.
Compensation of the global motion should be possible even if the observed scene is not planar. To be able to be used in a surveillance system, the compensation must be done in real time. Therefore, fully 3D scene reconstruction as done in structure from motion is not an option.
To keep the required processing power low, a full 3D scene reconstruction should be avoided. In video coding, the mesh-based motion compensation technology has proved to be able to compensate shape of objects better than normal block-based compensation. We adopted this approach to compensate global motion in a scene containing 3D objects.
