Stochastic Modeling of Camera Errors for Stereo Image Processing

Stereo vision algorithms with dense outputs are of great importance in the field of computer vision, virtual reality, environment reconstruction and robot navigation. In the field of robot navigation, mobile robots and fully automated vehicles have to be tracked in an environment they share with pedestrians. For this purpose, an increasing use of stereo cameras is expected in the future due to the significant improvements in the quality of distance measurements that can be achieved compared to mono cameras. The stochastic properties of such a sensor are crucial for the decision of the optimal tracking algorithm, e.g. Kalman Filter for a Gaussian probability density function. However, in the current research, mostly only the accuracy of the algorithm is analyzed. In this paper, it is tried to close the gap. The stochastic properties of a stereo image processing chain are analyzed. In experimental results, it is shown that the reprojection error is Gaussian. This contribution is very promising, as these results indicate that a Kalman Filter is the proper choice as a tracking algorithm in the image plane. Additionally, first results regarding the distance measurements achieved by the stereo image processing chain are presented. It is shown that the error from triangulation int this setup is Gaussian with an additional systematic error.