A high level approach to the interpretation of motion in dynamic scenes
Irwin King, Michael A. Arbib · 1993
Visual motion interpretation is a pervasive phenomenon vital for living organisms to survey, navigate, and interact with the environment. These behaviors require a rough sample measurement and the appropriate interpretation of the changing surroundings. Often when viewing interesting dynamic objects, the observer evaluates the structure by relying on the complex motion patterns created by the various sub-parts of the dynamic structure. The integrative theme of this dissertation is to propose a new approach to the representation and interpretation of these spatiotemporal texture patterns. We introduce the term Temporal Signatures to describe these unique high-level dynamic features. These special signatures are used to characterize rigid, articulated, and coordinated moving structures, e.g., biomechanical movements, over an extended space-time domain. Two significant and beneficial motion invariant properties for motion segmentation and grouping processes, Motion Reversal and Motion Symmetry, are also discussed in this framework. Furthermore, we offer a computational neural network approach which develops and implements perceptual neural mechanisms that are sensitive to these spatiotemporal patterns for motion segmentation and Temporal Signature extraction. In addition, an attention mechanism is implemented at this level to parse out the relative component motion in a moving object. Lastly, to achieve object recognition we demonstrate the matching of an input spatiotemporal structure to a dynamic object database using a labeled-graph method with the Dynamic Link Architecture. (Copies available exclusively from Micrographics Department, Doheny Library, USC, Los Angeles, CA 90089-0182.)