ARIS

(workpackage 1 - loria results)

 

The goal of the ARIS project was to provide new and innovative AR-technologies for e-(motion)-commerce application, where the products can be presented in the context of their future environment. Two system approaches have been developed:

 

1.      An interactive desktop system, where the end-user can easily integrate 3D product models (e.g. furniture) into a set of images of his real environment, taking consistent illumination of real and virtual objects into account.

 

  1. A mobile AR-unit, where 3D product models can be directly visualized on a real site and be discussed with remote participants, taking shared augmented technologies into account. Both systems have been implemented based on the same technology infrastructure and tested on well defined application scenarios with end-users.

 

Our contribution concerns real-time camera tracking. Several methods have been developed in order to obtain fast, accurate and robust tracking over long sequences :


markerless tracking using planar structures


During the first year of the project, we proposed a real time camera tracking system for scenes which contain planar structures. The type of scenes which can be considered with our method is large: this is commonly true for indoor environments where the ceiling or the ground plane are often visible. This is also often true for urban outdoor scenes because the façades of the buildings, the roads and the squares are often visible and can be used for registration.

 

The main idea of the algorithm is to compute the homographies induced by each visible plane. As the homography is a function  of the camera parameters,  the camera pose can be inferred from the set of homographies induced by the visible planes. This first prototype allows us to reach an approximate processing rate of 16 frames per second.

videos

x x
Indoor multiplanar tracking + augmentation
Outdoor multiplanar tracking + augmentation

papers

  1. Gilles Simon and Marie-Odile Berger. Pose Estimation for Planar Structures. IEEE Computer Graphics and Applications, 22(6):46-53, November 2002. [Abstract] [bibtex-entry]
  2. Gilles Simon and Marie-Odile Berger. Real time registration of known or recovered multi-planar structures : application to AR. In 13th British Machine Vision Conference 2002 - BMVC'2002, Cardiff, United Kingdom, pages 567-576, September 2002. [Abstract] [bibtex-entry]
  3. Gilles Simon and Marie-Odile Berger. Reconstructing while registering : a novel approach for markerless augmented reality. In International Symposium on Mixed and Augmented Reality - ISMAR'02, Darmstadt, Germany, September 2002. [Abstract] [bibtex-entry]


model selection


Even when the precision of the viewpoints is improved  by considering several planes, fluctuations in the parameters are often observed and may lead to unpleasant visual impressions such as jittering or sliding when augmented scenes are considered. These fluctuations are especially conspicuous when the camera motion is small because of noise and imprecision in computing the  points coordinates. In the past, several papers used Kalman filtering for prediction and stabilization task. However, the use of a Kalman filter is  not always advantageous for AR. This is because a low

order dynamical model of human motion  may not be always appropriate  except under very constraint scenarios.

 

Following Matsunaga and Kanatani [kanatani00] and [torr98] we investigate  the use of motion  model selection to reduce fluctuations of the camera parameters and to improve the visual impression of the augmented scene. The underlying idea in model selection is as follows: a higher order motion model fits any data set more accurately than a lower order model. However, high order models fit part of the random noise they are supposed to remove. Thus, a high order model, although accurate, is less stable to random perturbations in the data. A good motion model must strike the right balance between accuracy and stability.  The model selection principle demands that the model should explain the data very well and at the same time have a simple structure.

 

Within  the ARIS project, we decide to use together the model selection strategy and the multi-planar calibration in order to improve the stability and the accuracy of the estimated  the parameters. There are different branches using model selection, but there is no such successful criterion in general.  For this reason, we try to compare different model selection and we  especially consider the criteria which involve the covariance matrix on the estimated parameters and the Fisher information matrix. Indeed, often, criteria such as Akaike are only asymptotic approximations of a  criterion which includes the covariance or  the information matrix. So, we hope that such criteria will improve the model selection. 

 

As expected, the experiments which were conducted on synthetic images proved that the criteria which involved the covariance matrix or the Fisher information matrix gave the best results. The experiments using real image sequences taken with  a turntable also proved that the accuracy of the recovered camera pose is improved by model selection.  In addition, model selection produces smoother trajectory and better visual impression. For the closed sequence, Figure  1 exhibits the  distance from the current camera pose to the initial camera pose. The three curves respectively plot the actual camera pose, the pose recovered without model selection and the pose recovered when model selection is used.  This graphic proves that the use of model selection improves the accuracy of the viewpoint and reduces noticeably the drift problems that are common when long sequences are considered.

Figure 1: comparison of tracking performance with and without model selection.

 

Results on a miniature scene and on a real-size indoor sequence acquired using a hand held camera are shown below. In the second example, due to the brightness of the floor, some sheets of paper were put down on it to make easier the tracking process. During the sequence, two panoramic motions were realized, one with a tripod and the other  without  a tripod. Both are correctly labelled by the model selection process (The red cross indicates the stationary model, the green circle corresponds to the panoramic rotation, and the blue square to the general model).

 

To conclude, the use of model selection with various criteria proved that criterion involving information on the covariance of the estimated parameters are well suited to camera stabilization. They allow us to produce smoother trajectories and better visual impression. In addition,   model selection reduces noticeably the drift problems that are common when long sequences are considered.

 

[kanatani00] C. Matsunaga and K. Kanatani}. Calibration of a Moving Camera Using a Planar Pattern: Optimal Computation, Reliability Evaluation and Stabilization by Model Selection. In Proceedings of 6th European Conference on Computer Vision, Trinity College Dublin (Ireland), pages 595—609, 2000.

 

[torr98] P.H.S. Torr, A.W. Fitzgibbon and A. Zisserman. Maintaining Multiple Motion Model Hypotheses over Many Views to Recover Matching and Structure. In Proceedings of 6th International Conference on Computer Vision, Bombay (India), pages 485—491, 1998.

videos



Without model selection
With model selection
A real-size indoor sequence

papers

  1. Javier-Flavio Vigueras-Gomez, Marie-Odile Berger, and Gilles Simon. Iterative Multi-Planar Camera Calibration : Improving stability using Model Selection. In Eurographics Association, editor, Vision, Video and Graphics (VVG)'03, Bath, UK, July 2003. [Abstract] [bibtex-entry]


hybrid tracking


As pure vision-based tracking methods cannot usually keep up with fast or abrupt user movements, a hybrid approach which combines the accuracy of vision-based methods and  the robustness of  inertial tracking systems has been studied during the last year.

 

Some systems have been proposed in the past that combine vision and sensors. In [you01], an extended Kalman filter combines landmark tracking and inertial navigation. However, extended Kalman filters require good measurement models which are difficult to obtain in AR where the user is generally free of his motions.  More often, sensors are used as prediction devices to help image feature detection:  in [state96], a magnetic sensor is used to help landmark search, whereas in [klein02] an inertial sensor  is used for detecting edges corresponding to a wire-frame model of the scene.

 

Our approach is close to these  works, as we also use an inertial sensor to improve the matching stage. However, we brought significant improvements to these works:

 

Our marker-less tracking system is used as a basis of  the hybrid algorithm. The inertial sensor robustness allows  us to maintain tracking during long sequences. However, the process is  incremental and may progressively diverge because of successive approximations, or even stop when matching fails. 

 

For that reason, we finally proposed to add markers into the scene, whose positions are known. These markers are used to initialize or re-initialize the tracking process  in case of matching failure or divergence. Once the initial viewpoint is known, the hybrid system is used for tracking and the markers are not used anymore (they can disappear from the user field of view). As a result we get a system that is really robust against fast camera motions, but also very accurate (no jittering effect) as the whole texture information contained on the planar surfaces is used for tracking. Moreover, the user is free of his motions. It has been tested on real and miniature scenes and it runs at 10 fps on a 1.7 Ghz laptop.

 

Our implementation is based on an inertial sensor provided by XSens (model MT9). We accurately measured the sensor errors using a motorized pan/tilt unit. These errors are handled in the tracking process. In addition, we provided a tool  to perform the sensor/camera device calibration: indeed, to estimate camera rotations from sensor rotations, it is necessary to know the rigid transformation between these two devices.  Our tool is based on automatic detection of a calibration target for several positions of the sensor / camera device, and a classical hand-eye calibration is performed [tsai89].

 

[klein02] G. Klein and T. Drummond. Tightly Integrated Sensor Fusion for Robust Vision Tracking. In Proceedings  of the British Machine Vision Conference, BMVC 02, Cardiff, pages 787—796, 2002.

 

[state96] A. State, G. Hirota, D. Chen, W. Garett and M. Livingston. Superior Augmented Reality Registration by Integrating Landmark Tracking and Magnetic Tracking. In Computer Graphics (Proceedings Siggraph New Orleans), pages 429—438, 1996.

 

[tsai89] R. Y. Tsai and R. K. Lenz.A New technique for Fully Autonomous and Efficient 3D Robotics Hand/Eye Calibration. In IEEE Transactions on Robotics and Automation 5(3):345—358, June 1989.

 

[you01] S. You and  U. Neumann. Fusion of vision and gyro tracking for robust augmented reality registration. In Proc. IEEE Conference on Virtual Reality, pages 71—78, March 2001.

videos





Purely vision-based tracking with an abrupt motion
Same sequence using sensor prediction
A real-size sequence


Presentation of the complete mobile ARIS system

Ismar'04 demo : scene reconstruction, hybrid tracking and (to compare) tracking using markers only

papers

  1. Michael Aron, Gilles Simon and Marie-Odile Berger. Handling uncertain sensor data in vision-based camera tracking. In International Symposium on Mixed and Augmented Reality - ISMAR'04, Arlington, VA, November 2004.