Comparison of three methods for estimation of 3D object motion
This demonstration is on automatic analysis of moving objects in a video sequence. The goal is the estimation of the 3D motion between two time instants of a 3D rigid object.
Three methods are compared:
The first method assumes
and estimates 3D motion by an approach based on epipolar geometry. In the special case of orthogonal projection, only one of 6 3D rigid motion parameters can be estimated: The object rotation axis tilt that defines the orientation of the rotation axis projected into the image plane [LeSaux].
The experiments are carried out fully automatically. The point correspondences are estimated by a choice of reliable and spatially well distributed vectors from a dense displacement vector field. The 2D objects segmentation mask is derived based on change detection by the public software COST 211 Analysis Model [COST].
The second method assumes:
and estimates 3D motion by an optical flow approach. Using the known 3D shape, 5 from 6 3D rigid motion parameters are introduced in the optical flow equations. Due to the orthogonal projection assumption, the sixth parameter, the translation in viewing direction, can not be estimated [LeSaux].
The experiments are carried out fully automatically. The 2D objects segmentation mask is derived based on change detection by the public software COST 211 Analysis Model [COST]. A rough approximation to the 3D object shape is calculated from the 2D segmentation mask using a distance transform assuming convex shape.
The third method assumes:
and estimates fully automatically the object rotation axis tilt (see first method) by exploiting only object shading [Stau99].
The experimental results are derived fully automatically. The necessary input data for the algorithms is derived automatically too as follows:
|
||||
|---|---|---|---|---|
|
||||
|
The results visualize the projection of the 3D object rotation axis into the 2D image plane. The estimated motion follows qualitatively the motion of the object for all three methods. The results show that the applied assumptions are in general valid for simple video scenes.
However, the first method (based on point correspondences) shows the highest temporal stability. The second method (based on 3D shape) seems to fail for some images. The third method (based on shading) gives always slightly biased results due to the assumption that the main illumination is in viewing direction.