Tensor Space Learning for Analyzing Activity Patterns from Video Sequences

Discovering knowledge from video data has recently at- tracted growing interest from vision researchers. In this pa- per, we describe a tensor space representation for analyzing human activity patterns in monocular videos. Given a set of moving silhouettes derived from raw video data, the pro- posed methodology first learns a tensor subspace model to embed the silhouettes into low-dimensional projection tra- jectories with preserved temporal order. Symmetric mean Hausdorff distance is then used to measure dissimilarity be- tween the embedded motion trajectories in the tensor sub- space, as the basis for supervised or unsupervised learn- ing. The experimental results on two recent video data sets have shown that the proposed method can effectively ana- lyze human activities with intra- and inter-person variations on both temporal and spatial scales.