Translation-invariant motion perception for multiple objects using grid partitioning representation

An HMM-based motion perception system capable of identifying individual motion of multiple objects present in a scene has been developed. By introducing a new grid-partitioning scheme representing the characteristic motions in the scene, translation invariant perception has been established. An autonomous time indexing scheme has also been developed and implemented in the system. Recognition experiments were carried out for simple arm moving patterns in a variety of environments including translation and scaling of object images as well as the mixture of different patterns in a single scene. The average recognition accuracy of above 81% has been achieved in each experiments.