Self-organizing shape description for tracking and classifying multiple interacting objects

The problem faced in this work is related to tracking and recognition of rigid and non-rigid interacting objects in complex scenes from a static camera. The processing steps leading to the description of the behavior of objects in terms of trajectories and typology will be illustrated in details and the performances of the system will be discussed. The proposed approach uses an empty reference image for object extraction through image difference; the reference frame is updated continuously by a high level background updating module taking into account the detected objects and their classification tag. The tracking module is responsible for objects labelling being able to preserve objects identity even when an occlusion occurs on the image plane between different objects. A novel approach is considered for tracking and recognition, which is based on different features selection strategies applied to an initially redundant set of shape points (i.e. corners). Short-term and long-term memory models are used in a cooperative scheme. The two level feature selection strategy used by long-term shape models is described: at lower level a spatial-temporal voting method is used to assess temporal stability of spatial groups of corners; at the higher level, a supervised self-organizing scheme is used for objects classification.