Unsupervised Temporal Segmentation of Human Activities in Video

Human motion analysis from video has attracted much interest from computer vision researchers due to its promising applications for visual surveillance, perceptual user interface, content-based image storage and retrieval, video conferencing, athletic performance analysis and virtual reality. A key aspect to understand and build computational models of human actions is the capability to temporally segment actions in video. Several issues contribute to the challenge of temporal segmentation of human motion from video. These include the large variability in the temporal scale and periodicity of human actions, the complexity of representing articulated motion, and the exponential nature of all possible movement combinations. We formulate the temporal segmentation problem as an unsupervised learning problem, and propose Aligned Cluster Analysis (ACA), an extension of standard kernel k-means clustering to cluster time series. ACA extends standard kernel k-means clustering in two ways: (1) allows the cluster means contain a variable number of features and (2) introduces a generalized dynamic time warping (DTW) kernel as temporal metric between sequences. Experimental results reported on the Weizmann and KTH action datasets demonstrate the effectiveness of ACA for factorizing human actions in video.