Efficient local filter bank with over complete spatiotemporal pooling in action recognition

Action recognition task has been researched for years, algorithms based on local spatiotemporal interest points have gained successful results. However, these methods mainly face the problems like: The STIPs detectors only extract a sparse set of features and lack of structural orders. In this paper, we present a local motion filter bank with haar3D filters for action recognition. The filter bank is convoluted with the input volumes to decompose the motions as directions. Then an over complete spatiotemporal pooling stage is advocated to invariant to small shifts and hold the spatiotemporal information. Finally, sharing features are selected from the descriptors to form sparse linear models. The performance is tested in public data sets and gained reasonable results.