Trinocular vision
Method and system for real time high-resolution 3D images acquisition based on trinocular stereovision
Project founded by Romanian Ministry of Research (2004-2006)Introduction
The main disadvantage of stereo (2 camera) vision is that the disparities (differences between corresponding points in the two images) are, mainly, parallel with the baseline of the system (the line that joins the optical centers of the two cameras). The search for corresponding points thus needs to be done along lines parallel with this baseline. In the case of canonical or almost canonical stereo systems, these lines are horizontal (fig. 1).
Figure 1. Horizontal search line, detection of a vertical edge
To be able to precisely locate corresponding points, data with high frequency components needs to be present along horizontal lines. This high horizontal frequency corresponds to vertical (or at least non-horizontal) edges in the image. That means that horizontal edges are not easily detected (fig. 2).
Figure 2. Horizontal edges cannot be precisely located by a horizontal pair of cameras
The introduction of a third camera, such that it forms a vertical pair with the left camera, allows the correlation and 3D reconstruction of points laying both vertical and horizontal edges (fig. 3).
Figure3. Correlation of both vertical and horizontal edges: vertical edges are correlated using the left-right pair and horizontal edges using the left-bottom pair
Objectives
Development of a real-time trinocular system, capable of precisely reconstructing the 3D points from a scene, from images acquired by 3 synchronized cameras. Most of the edge points must be reconstructed, regardless of the orientation of the edge. The system must be able to reconstruct 3D points located in a wide range of distances. The reconstruction process must be as fast as possible in order to allow further processing to take place.
Achievements
A precise real-time trinocular reconstruction system was implemented, that allows edge-points of various orientations to be reconstructed.
Figure 4. Left: points reconstructed using a stereo system, and projected on the left image. Right: points reconstructed using the trinocular system, and projected on the left image. More horizontal edges are constructed.
a. The reconstructed scene
b. Binocular reconstruction-side viev
c.Trinocular reconstruction-side viev
d.Top views of the reconstructed points: with binocular system on the left and with trinocular system on the right
Figure 5.Image of a corridor and reconstructed points, using the binocular and the trinocular system, viewed from the side and the top. With the trinocular system the horizontal floor markings are precisely reconstructed.
a. Road scenario with 600m horizontal curvature
b. Binocular reconstruction-side viev
c.Trinocular reconstruction-side viev
d.Top views of the reconstructed points: with binocular system on the left and with trinocular system on the right
Figure 6. Image of a road scenario and reconstructed points, using the binocular and the trinocular system, viewed from the side and the top. With the binocular system horizontal and oblique features as the side lane delimiters are missed. With the trinocular system these features are reconstructed
Publications
S. Nedevschi, S. Bota, T. Marita, F. Oniga, C. Pocol, "Real-Time 3D Environment Reconstruction Using High Precision Trinocular Stereovision", 2006 IEEE-TTTC International Conference on Automation, Quality[[[[&]]]]Testing, Robotics AQTR 2006 (THETA 15), May 25-28 2006 Cluj-Napoca, Romania, Vol.2, pp. 333-338.
S. Bota, S. Nedevschi, "GLSCENEINT: A Synthetic image generation tool for testing computer vision systems", Poster volume of IEEE 2nd International Conference on Intelligent Computer Communication and Processing, 1-2 September, Cluj-Napoca, Romania, 2006, pp. 39-44
S. Bota, S. Nedevschi, "Real-time trinocular vision for driving assistance systems", Proceedings of IEEE 2nd International Conference on Intelligent Computer Communication and Processing, 1-2 September, Cluj-Napoca, Romania, 2006, pp. 69-76