Motion stereo at sea: Dense 3D reconstruction from image sequences monitoring conveyor systems on board fishing vessels

A system that reconstructs 3D models from a single camera monitoring fish transported on conveyor is investigated. Models are subsequently used for training species classifier and improving estimates of discarded biomass. It demonstrated monocular camera, combined with conveyor's linear motion produces constrained form multiview structure motion, allows the scene to be reconstructed using conventional stereo pipeline analogous binocular camera. Although was proposed several decades ago, present work first compare accuracy precision cameras conveyors operationally deploy system. The exploits Convolutional Neural Networks (CNNs) foreground segmentation matching. Results laboratory model show when mounted 750 mm above conveyor, median <5 can achieved an equivalent baseline 62 mm. largely limited by error in determining (i.e. distance travelled belt). When ArUco markers placed belt, inter quartile range (IQR) z (depth) near optical centre found ±4