Foreground Segmentation Using Motion Vectors in Sports Video

Abstract

In this paper, we present an effective algorithm for foreground objects segmentation for sports video. This algorithm consists of three steps: low-level features extraction, camera motion estimate, and foreground object extraction. We employ a robust M-estimator to motion vectors fields to estimate global camera motion parameters based on a four-parameter camera motion model, followed by outliers analysis using robust weights instead of the residuals to extract foreground objects. Based on the fact that foreground objects’ motion patterns are independent of the global motion model caused by camera motions such as pan, tilt, and zooming, we considers those macro-blocks as foreground, which corresponds to the outliers blocks during robust regression procedure. Experiments showed that the proposed algorithm can robustly extract foreground objects like tennis players and estimate camera motion parameters. Based on these results, high-level semantic video indexing such as event detection and sports video structure analysis can be greatly facilitated. Furthermore, basing the algorithm on compressed domain features can achieve great saving in computation.