Automatic acquisition of movement information by a knowledge-based recognition approach

Abstract

We discuss automatic acquisition of movement information from two types of input data using a knowledge-based scheme. When the input data is an image of 3D surface points mensurated from a single perspective, the process of recognizing and locating a body position in each image frame includes (1) isolating the body from the motion scene, (2) segmenting it into the individual body parts, and (3) identifying these parts with the labeled body parts of a stored model. When the input data is a limited number of 3D points indicating key body parts in activities, only step (3) is required for movement recognition: we only need to recognize light emitting devices (LED's) attached to the reference points without going through every detail of body surface points. The constraints in the system are explored and propagated to extract structural information from the input data compatible with the model structure.