Multi-view Human Action Recognition by Cell Summary Descriptor and Decision Fusion

Multi-view human action recognition (MVHAR) is essential for many applications including smart video surveillance in shopping malls, airports, railway stations and other public places, as well as the ambient assisted living systems. Currently, a lot of methods realized MVHAR by constructing high dimensional features or by complex calculation process, which causes the recognition speed difficult to meet the needs of real-time application system. To address the problem, a cell summary (CS) descriptor is constructed for each frame by dividing each frame into several cells and further dividing each cell into several radial bins and then extracting spatiotemporal features from each radial bin. Then a video is represented by a sequence of dimension-reduced CS descriptors. The CS descriptor is not only discriminative but also low computational cost. A probabilistic classifier is learned for each view of each action category, and then action classification is carried out independently in each view. A probability estimation based decision fusion algorithm is proposed to make a final decision. Experimental results on the two publically available multi-view human action datasets MuHAVi-MAS-14 and IXMAS show that the proposed approach is superior to state-of-the-art methods in recognition accuracy and moreover, the approach is so computationally efficient that it is appropriate for real-time applications.