A Scale-Aware Monocular Odometry for Fishnet Inspection With Both Repeated and Weak Features

Remote operated vehicles (ROVs) carrying various sensors can regularly perform inspection tasks in fishnet areas, hence reducing manpower and financial consumption. However, the repeated texture features of the fishnet cause perceptual ambiguity in the state-of-the-art stereo vision technologies, and monocular vision methods can only provide up-to-scale pose estimation results. This article proposes a scale-aware monocular odometry based on direct sparse method (DSO) to yield accurate estimation on vehicle’s position for fishnet inspection. Especially, a region-of-interest (ROI) extraction method is proposed to extract fishnet images containing only repeated textures, enabling the odometry initialization to be completed in any region of the cage. Moreover, absolute scale is obtained by aligning the mesh model with the real fishnet, where the mesh is parameterized onto the tangent plane. The proposed method is experimentally evaluated on a sea trail using an ROV in Sanya, China. Experimental results show that our method can yield accurate odometry estimate for underwater vehicles even with both repeated and weak features in fishnet.