3D object motion from matching points

Estimation of 3D object motion two succeeding video images with matching points

Goal

This demonstration is on automatic analysis of moving objects in a video sequence. The goal is the estimation of the 3D motion of a 3D object.

Approach

The 3D object motion is estimated using the following assumptions:

• 2D object segmentation mask is known
• Object is rigid
• Scene illumination is diffuse
• Camera can be approximated by orthogonal projection

The object motion is estimated by an epipolar geometry approach. The tilt angle of the object rotation axis is extracted from epipolar equations which link points correspondences. The 2D object segmentation mask is used to select point correspondences belonging to the object. [LeSaux].

Experimental Results

The experimental results are derived as follows:

• Object segmentation: By COST 211 Analysis Model [COST]
• 3D motion estimation: For each image w.r.t. the preceding image

Original image sequence Object segmentation mask Estimated projected rotation axis

The results visualise the projection of the 3D object rotation axis into the 2D image plane. Apart of the beginning of the sequence, where the object does not rotate, the estimated axis follows well the motion of the object. The results show that the applied assumptions are valid for simple video scenes.

References

[COST]

COST 211quat Analysis Model

[LeSaux99]

B. Le Saux: "Estimation du mouvement d'un objet avec un modèle de caméra simplifié", rapport de stage de DEA, IRISA/INRIA, Rennes, France, Septembre 1999.

Up to motion page