Gender Recognition from Walking Movements using Adaptive Three-Mode PCA

We present an adaptive three-mode PCA framework for recognizing gender from walking movements. Prototype female and male walkers are initially decomposed into a sub-space of their three-mode components (posture, time, gender). We then assign an importance weight to each motion trajectory in the sub-space and have the model automatically learn the weight values (key features) from labeled training data. We present experiments of recognizing physical (actual) and perceived (from perceptual experiments) gender for 40 walkers. The model demonstrates greater than 90% recognition for both contexts and shows greater flexibility than standard PCA.