Multi-model Switching sequential image stitching for ROV-based visual mapping of offshore aquaculture netting

Image stitching can extend the limited field of view of ROV imagery and provide continuous visual records for aquaculture netting inspection. However, the repetitive, weak-texture characteristics of netting and deformations induced by water currents make it difficult for conventional stitching methods to generate complete panoramas. This paper proposes a factor graph-based sequential image stitching framework for visual mapping of aquaculture netting. Specifically, a motion prior-based image registration algorithm incorporating a planar netting model is introduced to improve the reliability of image correspondences. Furthermore, a multi-model switching framework uses the spatial distribution and tracking quality of optical-flow correspondences to trigger covariance inflation of the optical-flow constraint and introduce ORB and odometry constraints for homography estimation. The proposed framework is evaluated using simulated, controlled land-based, and sea-trial datasets. The results show robust and efficient sequential stitching in netting scenes. The resulting panoramic visual records provide a basis for ROV-assisted inspection of offshore aquaculture cages. • Motion priors reduce false matches in repetitive, weak-texture netting. • A switching factor graph fuses optical flow, ORB features, and odometry. • Score-triggered covariance inflation downweights unreliable optical-flow constraints. • Efficient ROV netting stitching is tested in simulation, on land, and at sea.