Robust Scene Interpretation of Underwater Image Sequences

A machine vision system is proven at a conceptual level to help unmanned remotely operated vehicles (ROV's) interpret underwater oceanic scenes and clarify poor quality image sequences. The novel image processing approach is both data-driven and expectation-driven by combining an image with its predicted version. The system maintains an interpretable display to the ROV operator during large variations in image quality. Information displayed is enhanced with a clearer computer model to enable the operator to recognise the ROV's position relative to an underwater oil-rig. Images contain objects of interest (metal cylinders of a known oil-rig structure) and background (water). After linear contrast-stretching for enhancement, images are segmented using a method based on a Markov Random Field (MRF) model combined with a rapidly convergent deterministic technique termed iterated conditional modes. Cylinders are analysed to determine the camera's viewing direction, range and twist off the vertical. The camera's position is then calculated given knowledge of the node in view extracted from 3D model data. Successive viewpoints from a sequence of images are fed through a Kalman filter to predict the next viewpoint. Placing this in a 3D computer model of the structure allows a 2D predicted image to be projected. This prediction, combined with the next acquired image, improves image segmentation and subsequent scene interpretation using an extended MRF model. 'Proof of concept' results are demonstrated from frame-grabbed images sequences taken in air and through water. Laboratory sequences in air are degraded using software. A laboratory sequence is also tried through increasingly murky water. The work extends to include underwater lake trials using a mini-ROV. Video footage was frame-grabbed into an image sequence and processed off-line. Results show that the novel predictive feedback method is much more resilient to image degradation, successfully interpreting poor quality images that would otherwise go unrecognised