Depth-from-motion estimation based on a multiscale analysis of motion constraint equation validity

We present a new method of estimating the depth-from-motion in a image sequence. It is assumed that the camera motion is known and the scene is rigid. Our idea is to evaluate carefully the pixels where the brightness change constraint equation is valid, and consequently the depth value is reliable. A confidence measure is derived from local analysis. This analysis is performed at various levels of resolution, increasing the number of reliable pixels. The depth values obtained at each level are then fused within a statistical framework, allowing one to also integrate the temporal information. Results are shown for real and synthetic images used in the performance evaluation.