Skip to Main Content
Understanding human gestures can be posed as a typical classification problem. Within the computer, gestures are represented as time-varying patterns in feature space. These patterns, though variable, are distinct and have associated meanings. In the absence of a priori knowledge of the underlying class probabilities, classification is performed based on some notion of similarity, e.g. distance, among samples. The k-nearest neighbour (kNN) decision rule has often been used in these pattern recognition problems. The use of this particular technique gives rise to multiple issues, one of them being that it operates under the implicit assumption that all features are of equal importance in deciding the class membership of the pattern to be classified, regardless of their "relevancy". This paper presents an accelerometer-based gesture recognition system that utilizes Mahalanobis distance metric learning to derive optimal weighting scheme for nearest neighbour classification. The metric is trained with the goal of separating different classes by large local margins and pulling closer together samples from the same class, based on using as few features as possible. Our experiments on an arbitrary gesture set show that the proposed method leads to significant improvements in recognition accuracies, yielding simultaneously a maximum of feature discrimination.