Towards fast, view-invariant human action recognition

In this paper, we propose a fast method to recognize human actions which accounts for intra-class variability in the way an action is performed. We propose the use of a low dimensional feature vector which consists of (a) the projections of the width profile of the actor on to an ldquoaction basisrdquo and (b) simple spatio-temporal features. The action basis is built using eigenanalysis of walking sequences of different people. Given the limited amount of training data, Dynamic Time Warping (DTW) is used to perform recognition. We propose the use of the average-template with multiple features, first used in speech recognition, to better capture the intra-class variations for each action. We demonstrate the efficacy of this algorithm using our low dimensional feature to robustly recognize human actions. Furthermore, we show that view-invariant recognition can be performed by using a simple data fusion of two orthogonal views. For the actions that are still confusable, a temporal discriminative weighting scheme is used to distinguish between them. The effectiveness of our method is demonstrated by conducting experiments on the multi-view IXMAS dataset of persons performing various actions.