Decision Making Under Uncertainty in the context of Gesture and Trajectory Classification

Abstract

We aim to construct a classier of time-varying gestures, a project motivated by burgeoning research in video classification building upon lessons learned from successful large scale image classification. With an understanding of current techniques for static image classification, we built and experimented with static image classifiers for the gestures of rock,paper, and scissors, one using a support vector machine and another using a convolutional neural network. We then applied this learning toward the problem of classifying a trajectory tracked from a time-varying gesture video. We tracked a moving gesture target, and trained a support vector machine classifier with the resulting path. We also approached the problem of classifying a trajectory early using multiple support vector machines at even intervals and a 2 stage method of detecting primitive gestures and classifying compound trajectories using a grammar composed of primitives. These methods were tested against real world noise and trajectory variation situations to evaluate performance and robustness.