M. Grundmann, F. Meier, I. Essa
3D Shape Context and Distance Transform for Action Recognition

We propose a new method for human action classification from video sequences that uses 3D Shape Context to capture the spatial and temporal details. We generalize the 2D Shape Context and 2D Distance Transform approaches for representing and classifying actions. An action is modeled by a 3D point cloud extracted by sampling its silhouettes over time. We propose a new non-uniform sampling method, which gives preference to fast moving body parts and use the Euclidean 3D Distance Transform to determine these parts. Actions are classified by matching their point clouds based on 3D Shape Context, which is highly discriminative. Our approach is based on global matching and therefore improves on traditional local feature matching.
We test the approach thoroughly on two publicly available datasets and compare to several well-known and the state of-the-art methods. The achieved classification accuracy of our approach is on par with or superior to the best results reported to date. In addition, our proposed method is fast and does not need a prior learning stage.

Bibtex entry

not yet