Multi-object segmentation by stereo mismatch

Abstract

A new stereo mismatch based foreground object segmentation method is described. It efficiently locates objects over a wide range of depths against backgrounds of known 3D geometry, even in the presence of rapidly changing lighting and dynamic textures, such as projected video. Not relying on full stereo reconstruction, it is fast enough in software for some real-time applications, robust to camera quality, and requires little parameter tuning. Experimental results validate the approach, demonstrating its ability to simultaneously distinguish multiple objects in a complex scene, even when close together or partially occluded.