Calibration of Vision-Sensor Networks

by Kyng min Han

Automatic camera calibration of a multi-camera rig

The goal of this research is to eliminate any human intervention from calibrating multi-camera system. For a given image data set, the system finds the best subset of images which optimizes the camera calibration. The algorithm procedure is as following

1. Collect images: In the beginning, human just wander around inside of the camera rig. Each camera grab images at the same time.

2. Feature point detection: We use Hough transform to detect lines. Feature points are the intersection of these lines.

3. Selecting the best input data set: The proposed frame work determines the best subset of images that can optimize the camera calibration process (9 sets for each cameras)

4. Determining a reference coordinate frame: Once the calibration is done, then the algorithm decides the reference coordinate frame and a camera path tree. Using this path tree, one can easily bring image points in some other camera cameras to the reference frame.

5. Extrinsic parameter test: Above figure illustrates the real camera positions in our lab (left) and reconstructed camera frames (right) by their extrinsic parameters.

6. 3D Object reconstruction: Here I made a cross shape like object. The four corner points (Tips of each bar) are reconstructed.

Object views from 6 different cameras

Reconstructed points (red circles)

Results

a) Real Camera positions b) Reconstructed camera coordinate frames

References

1.  Han, K., Dong Y. and DeSouza, G.N., "Autonomous Calibration of a Camera Rig on a Vision Sensor Network," in the Proceedings of the 2010 IFAC International Conference on Informatics in Control, Automation and Robotics (ICINCO), June 2010, Portugal.
    

2. Han K., DeSouza, G. N., "A Feature Detection Algorithm for Autonomous Camera Calibration", in the Proceedings of the 2007 IFAC International Conference on Informatics in Control, Automation and Robotics (ICINCO), pp. 286-291, May 2007, France.