Body region labelling for action recognition

We present a novel method for automatically labelling the head, torso, and legs of a human body tracked through a video sequence. An appearance-based body model is constructed by dividing the initial silhouette into a series of spatial slices, and building a colour distribution histogram for each. In subsequent frames a labelling hypothesis is constructed for each new silhouette by matching against these distributions, and used to identify each body region under a range of poses. We use the body model to extract feature points, which we use as the basis for an action recognition scheme. Actions are represented by a vector of the head and torso positions, sampled over the duration of an action. Manually labelled sequences provide a training set comprising sitting, bending, squatting, and lying actions, viewed from various angles. We use nearest-neighbour matching to identify actions presented in test sequences. Our results show that our method is effective, achieving a high recognition rate.