| |
|
|
|
|
 |
Dieu Sang Ly |
 
 |
| Project:
|
| |
| Supervisors:
|
| |
| Host Universities:
|
| |
| Personal introduction |
top |
| I finished my undergraduate degree in Automatic Control in February 2006. After working as an electronic engineer for 5 months in a Japanese company, I joined the VIBOT Master program. Computer Vision is completely new for me as I had only electronic and automatic engineering background. I gained much interesting knowledge in Vision after this 2-year course. Especially I had chance to reside in 3 European countries and discover different cultures with classmates of 15 nationalities. |
|
| |
| Project Description and
Objectives |
top |
The omnidirectional vision systems enlarge the field of view of
conventional cameras by means of special lenses, multi-image acquisition systems or catadioptric cameras. The catadioptric cameras are
the customary ones coupled with mirrors of special curvatures to obtain an entire field of view. Structured light projection is widely
used to solve the correspondence problem in stereo computation. This thesis is focused on a structured light-based stereo system which
is formed by a catadioptric camera and an omnidirectional laser projector.
Calibration is a critical process before using a camera for any further purpose. Therefore over the past decades, many researches have
been dedicated to calibration methods for omnidirectional vision sensors in order to achieve efficient approaches.
This thesis is carried out at MIS laboratory, Amiens in order to obtain the Erasmus Mundus Master degree in Computer Vision and Robotics
(VIBOT). The objective is to study and improve the calibration of the structured light-based catadioptric stereo sensor. In the most
recent work on the calibration method for this sensor, the catadioptric camera and the omnidirectional laser projector are calibrated
sequentially to estimate the parameters of each component. This thesis proposes a novel calibration which permits a simultaneous
calibration for both of them. The proposed approach simplifies the process of sequential calibration while preserving the accuracy of
the scene reconstruction. Particularly the simultaneous approach provides an important improvement when it does not require an accurate
measurement of the calibration patterns in 3D space as the sequential approach does. After the sensor parameters are estimated, the
sensor model can be used for depth measurement which has many further applications such as scene reconstruction, robot navigation, etc.
|
|
| |
|
Since this work is based on the real prototype already developed in the laboratory (see figure 1), the first step is to review the arrangement, calibration and applications that have been accomplished for this prototype. Next, a novel calibration approach is developed to simplify the previous one and to enhance the subsequent scene measurement. A complete model of the sensor is obtained by using the projection model of the camera and the geometry of the projected laser pattern (see figure 2). The sensor parameters are then estimated using the complete model by a non-linear iterative algorithm such as Levenberg-Marquardt. The performance of the proposed method is evaluated by simulation. Once the new approach is validated with the synthetic data, the following step is to apply this method to the omnidirectional stereo sensor. Lastly, the calibrated system is used for scene reconstruction.
Figure 1 |
Figure 2 |
|
| |
|
The sequential and simultaneous calibration methods are evaluated and compared by simulation with a set of synthetic 3D and 2D points. The simulation results have shown that:
1. Concerning the sensor modelling, the sequential method uses the projection model of the camera while the simultaneous one adds to that model the representation of the laser projection. However, the laser projection is easily modelled by geometry; hence, there is no significant complexity.
2. Two calibration methods provide similar reconstruction accuracy if the 3D points use for the calibration are well known. The following figures illustrate the reconstruction errors of the sequential and simultaneous calibrations with a surrounding environment of small range (figure 3) and large range (figure 4).
Figure 3 |
Figure 4 |
3. The sequential method requires the calibration patterns placed around the sensor whereas the simultaneous one does not as it can employ directly the laser profile projected to the scene. Moreover, in the sequential approach, the 3D points should be well measured to ensure a reliable calibration. This limitation is overcome by the simultaneous calibration in which the 3D laser points can be approximated from the structure of the scene. The next figure shows the comparison result of two approaches with the increased error in the approximation of the 3D points. It is obvious that the simultaneous method is much better than the sequential one if the 3D points used for calibration are not accurately measured.
Figure 5 |
|
| |
| First of all, I would like to thank Mr. El Mustapha Mouaddib for sharing with me his experience, thorough instruction and passion for research. I should show my gratitude to the teaching staffs of the VIBOT Master. I will never forget VIBOT friends and the great time we had together. Finally, I want to express my deep gratitude and love to my family. |
|
| |
|
|
| |
|
|
|
|
|
|
|
|
|
|
|
