Reconstructing force-dynamic models from video sequences

摘要

This paper presents a method for recovering the support, contact, and attachment (i.e., force dynamic) relations between objects depicted in video sequences. It first presents a stability-analysis procedure that determines whether a configuration of observed objects is stable under a given interpretation using a reduction to linear programming. It then presents a model-reconstruction procedure for searching the space of admissible interpretations to find the simplest stable interpretations or models. These stability-analysis and model-reconstruction procedures have been implemented as part of a system that recovers force-dynamic interpretations from video sequences and uses those interpretations to classify the events that occur in those sequences. This paper presents the details of the stability-analysis and model-reconstruction procedures and illustrates their operation on sample video sequences.