• Summary
• Related Documents
  • Active Bibliography
  • Co-citation
• Version History

Abstract:

Many autonomous vehicles are equipped with a camera based sensor to retrieve information about its environment. One important issue when dealing with mobile systems is that egomotion of the camera can seriously complicate the interpretation of the observations. Egomotion introduces additional motion components in the observed motion in the image plane. To enable accurate interpretation of the recordings of a camera, these 'disturbances ' must be taken into account. This paper presents an approach for estimating these disturbances. We divide the egomotion problem in 2 parts: we separate the direction of the translational motion of the camera (the mean viewing direction, measured relative to the driving direction of the vehicle) from its small inter-frame rotations. The dynamical characteristics of these two parts differ a lot, asking for different estimation approaches. The estimation of the mean viewing direction is directly related to the estimation of the location of the so-called Focus-of-Expansion (FOE). We estimate the location of the FOE based on a minimal number of 2 tracked background points. Confidence measures are used to ensure the reliability of the estimation. The inter-frame rotations are estimated from the complete optical flow field of background points. We use a (re-weighted) orthogonal least squares technique to estimate the small rotational components from the flow field data. Experiments based on practical video data illustrate the performance of our method in practice. The performance of the different orthogonal least squares techniques used to estimate the inter-frame rotations are compared. The results indicate the robustness and high accuracy of our approach for estimating both the mean viewing direction of the camera and its inter-frame rotations.

Citations