Human Action Retrieval via Spatio-temporal Cuboids

An approach for human action retrieval in videos is proposed. Based on the volumetric analysis, actions in videos are represented by part-based cuboids. To make full use of the structural information, an explicit shape model (ESM) is designed for probabilistic latent semantic analysis (pLSA). To ensure enough cuboids can be extracted, an improved detector is used at multiple frequencies. Experimental results on KTH dataset validate that our approach enhances the performance of pLSA, and results on surveillance videos prove it can deal with multiple actions well.