Learning non-negative locality-constrained Linear Coding for human action recognition

Description methods based on interest points and Bag-of-Words (BOW) model have gained remarkable success in human action recognition. Despite their popularity, the existing interest point detectors always come with high computational complexity and lose their power when camera is moving. Additionally, vector quantization procedure in BOW model ignores the relationship between bases and is always with large reconstruction errors. In this paper, a spatio-temporal interest point detector based on flow vorticity is used, which can not only suppress most effects of camera motion but also provide prominent interest points around key positions of the moving foreground. Besides, by combining non-negativity constraints of patterns and average pooling function, a Non-negative Locality-constrained Linear Coding (NLLC) model is introduced into action recognition to provide better features representation than the traditional BOW model. Experimental results on two widely used action datasets demonstrate the effectiveness of the proposed approach.