This is now an inactive research group it's members have moved on. You can find them at their new research groups:

Research Overview

[Model image]
Click on the above image for animation
Our approaches to model-based recognition first considered human motion to be that of a moving pendulum. This was also used to find subjects with articulated motion. It further allowed deployment of a gait signature based on the spectrum of the variation in inclination of the human thigh. Essentially, gait was modelled as Simple Harmonic Motion (SHM) and the cue to identity was the difference between perceived motion and that of pure SHM. As our particular focus was the variation in inclination of the thigh, we have also studied the nature of its variation, and developed measures that were invariant to trajectory and derived from the data itself. Recently, we have moved to using a continuous model formulation, based on our other work in feature extraction, leading to the notion that recognition can be based on motion templates. More recently, and especially for generality, we have deployed our evidence gathering techniques for 3D based recognition which also allows possibility of gait-motion capture via a marker-less system.

[UCSD Walking subject]
Click on the above image for animation
Our statistical approaches started from the notion that these approaches have proven application potential in automatic gait recognition. We first deployed Principal Components Analysis for data compression, coupled with Canonical Analysis for improved (and practicable) recognition, processing thresholded-image and optical flow data. Though this appears to be very promising, it could be reinforced by relationship to the mechanics of gait. This has been achieved by a new approach that uses velocity moments (statistical moments computed for moving objects).
© School of Electronics and Computer Science of the University of Southampton