Estimating Camera Tilt from Motion without Tracking

Most methods for automatic estimation of external camera parameters (e.g., tilt angle) from deployed cameras are based on vanishing points. This requires that specific static scene features, e.g., sets of parallel lines, be present and reliably detected, and this is not always possible. An alternative is to use properties of the motion field computed over multiple frames. However, methods reported to date make strong assumptions about the nature of objects and motions in the scene, and often depend on feature tracking, which can be computationally intensive and unreliable. In this paper, we propose a novel motion-based approach for recovering camera tilt that does not require tracking. Our method assumes that motion statistics in the scene are stationary over the ground plane, so that statistical variation in image speed with vertical position in the image can be attributed to projection. The tilt angle is then estimated iteratively by nulling the variance in rectified speed explained by the vertical image coordinate. The method does not require tracking or learning and can therefore be applied without modification to diverse scene conditions. The algorithm is evaluated on four diverse datasets and found to outperform three alternative state-of-the-art methods.