Local non-negative component representation for human action recognition

The Bag of Words (BOW) method with spatio-temporal interest points has achieved great performance in human action recognition. However the traditional BOW methods based on vector quantization (VQ) suffer serious quantization error and lose masses of information. There are two main reasons leading these: the first is the codebook obtained by k-means has no obvious visual interpretation and second, each input data is combined with only one label. In this paper, we apply non-negative matrix factorization (NMF) to learn codebook for actions, which provides intuitive non-negative components for actions. And then, Locality-constrained linear coding (LLC) method is applied to get the parts-based encodings for videos, which greatly alleviates the quantization error and considers the locality among bases and input samples. Our method is verified on the challenging database (KTH) and achieves commendable result.