Video Segmentation

Given a video $\mathbf{V}$, we first slice it into $N$ short segments: $\mathbf{V} = (v_1, v_2,...,v_N)$. Ideally, each segment $v_j$ would contain a single activity event. For simplicity, we slice the video into fixed time duration (4 seconds) with overlapping time window. The segmentation is not perfect, but the video segments typically contain enough information for determining the activity type, e.g. in the nursing home video, in 4 seconds people can take a few steps or pick up an object.