Analysis of Videos using Tile Mining

We investigate how mining top-k largest tiles in a data stream under the sliding window model can be useful for (real-time) analysis of videos and, in particular, for tracking. We first explain how a tracking problem can be cast into a stream pattern mining problem. We then show some preliminary results on tracking in the particular context where both the objects and the camera are moving and where the user does not specify the regions of interest in the first frames of the videos.