Detection and Recognition of Periodic, Nonrigid Motion

  • Authors:

  • Ramprasad Polana;Randal C. Nelson

  • Affiliations:

  • MicroStrategy Inc., 8000 Towers Crescent Drive, Vienna, Virginia 22182. E-mail: polana@strategy.com;Department of Computer Science, University of Rochester, Rochester, New York 14627. E-mail: nelson@cs.rochester.edu

  • Venue:
  • International Journal of Computer Vision
  • Year:
  • 1997

Quantified Score

Hi-index 0.00

Visualization

Abstract

The recognition of nonrigid motion, particularly that arising fromhuman movement (and by extension from the locomotory activity ofanimals) has typically made use of high-level parametric modelsrepresenting the various body parts (legs, arms, trunk, head etc.) andtheir connections to each other. Such model-based recognition hasbeen successful in some cases; however, the methods are oftendifficult to apply to real-world scenes, and are severely limited intheir generalizability. The first problem arises from the difficultyof acquiring and tracking the requisite model parts, usually specificjoints such as knees, elbows or ankles. This generally requires someprior high-level understanding and segmentation of the scene, orinitialization by a human operator. The second problem, withgeneralization, is due to the fact that the human model is not muchgood for dogs or birds, and for each new type of motion, a new modelmust be hand-crafted. In this paper, we show that the recognition ofhuman or animal locomotion, and, in fact, any repetitive activity canbe done using low-level, non-parametric representations. Such anapproach has the advantage that the same underlying representation isused for all examples, and no individual tailoring of models or priorscene understanding is required. We show in particular, thatrepetitive motion is such a strong cue, that the moving actor can besegmented, normalized spatially and temporally, and recognized bymatching against a spatio-temporal template of motion features. Wehave implemented a real-time system that can recognize and classifyrepetitive motion activities in normal gray-scale image sequences.Results on a number of real-world sequences are described.