There is a large and growing demand across many industries for underwater inspection technologies. Whether the assets in question are offshore oil rigs owned by energy companies, undersea power or data cables owned by utilities, or large pipe networks owned by municipalities, the problem is the same: critical assets are in need of assessment and maintenance but inspection is difficult and expensive due to their underwater location. Currently, much of the work in the industry consists of flying a remotely operated underwater vehicle equipped with a camera to take video of these assets in order to determine their state and whether maintenance is required. These videos can be many hours in length, and must be reviewed by a specialist at significant cost to assess the state of the asset.
This research project was undertaken in participation with 2G Robotics Inc., a company focused on underwater inspection. The goal is to implement computer software that can convert pre-existing single-camera videos of an asset, with unknown or varying camera parameters (auto-focus or zooming is allowed), to a photo-realistic 3D model of the asset. With such a model, the client can rotate, zoom, and generally inspect the asset much more quickly and intuitively without the need to review hours of video. There is demand for such software, and the resulting library could be licensed to inspection companies or offered under the Software-as-Service model.
Videos are reconstructed in three stages:
1) Video Tracking - image features are detected in the video, and tracked using optical flow. This produces a set of image correspondences, representing distinct 3D points imaged from many different locations.
2) Projective Reconstruction - using three-view geometry, the 3D structure of the scene and the motion of the video camera are solved simultaneously. The algorithm is robust to noise and mislabeled data.
3) Autocalibration - A projective reconstruction is not suitable for visualization or measurement - in projective geometry, concepts of distance and parallelism have no meaning. In order to transform the projective reconstruction into a metric space, the camera properties must be estimated. This research project has developed a hybrid autocalibration method that combines the best aspects of two existing algorithms. By making basic assumptions about the physical properties of the cameras, the hybrid method can produce a metric 3D reconstruction. In its current state, the 3D reconstruction algorithm can produce point cloud data from underwater video sequences. Future work will allow the point cloud to be meshed and textured with image data, producing a photorealistic 3D model.
Cavan, N., "Reconstruction of 3D Points From Uncalibrated Underwater Video", Systems Design Engineering, Waterloo, ON, University of Waterloo, 2011. Details