Homography-Based 3D Scene Analysis of Video Sequences *
Mei Han, Takeo Kanade · 1999
We propose a framework to recover projective depth based on image homography and discuss its application to scene analysis of video sequences. We describe a robust homography algorithm which incorporates contrast/brightness adjustment and robust estimation into image registration. We present a camera motion solver to obtain the egomotion and the real/virtual plane position from homography. We then apply the LevenburgMarquardt method to generate a dense projective depth map. We also discuss temporal integration over video sequences. Finally we present the results of applying the homography-based video analysis to motion detection. 1 Introduction Temporal information redundancy of video sequences allows us to use efficient, incremental methods which perform temporal integration of information for gradual refinement. Approaches handling 3D scene analysis of video sequences with camera motion can be classified into two categories: algorithms which use 2D transformation or model fitting, ...