The detection of moving objects in the environment is important for a broad range of tasks involving activities such as locomotion, manipulation, and description. This proposal describes a line of research aimed at studying vision-based techniques for recognizing the presence of moving objects in the field of view. The solutions must have broad generality and good reliability; the analysis specifically deals with problems associated with moving-object detection when the camera may also be moving. In the latter case, moving and stationary objects may both be changing position with respect to the camera, generating visual motion in the imagery. Significant preliminary work has been completed on the problem. The proposed work would build on the prior results by addressing three problems that are central to the development of a general understanding of detection of moving objects: (1) how can depth information be combined with visual motion? (2) can existing structure-from-motion algorithms be extended to include the detection of motion while avoiding problems with numerical instability? (3) can detectability thresholds be computed based on magnitudes of relative motion and noise, and how do various simplifying assumptions affect those thresholds?
