3D measurement of dynamic structures using monocular camera

Visual simultaneous localization and mapping technology assumes that the surrounding objects are stationary, making it inapplicable to dynamic objects. In environments with dynamic objects, accurately estimating their shape and displacement remains challenging. This paper develops a method for estimating the 3D shapes and displacements of dynamic objects using a monocular camera and demonstrates its effectiveness. The proposed technique uses semi-supervised video segmentation to separate dynamic objects of interest from others, extract and track features only for the object, and perform 3D reconstruction. Reconstructing these 3D point clouds in the camera space makes it possible to measure the 3D shape and displacement of a dynamic object. The method was validated on a video of a rotating fan blade, and the average relative absolute error of the position estimation was 0.0011, indicating sufficiently high accuracy. Hence, this approach highlights the possibility of accurately reconstructing dynamic environments using a monocular camera for infrastructure monitoring. • A method for 3D shape estimation of dynamic objects by a single camera is proposed. • Video object segmentation is used to separate dynamic objects of interest. • The method estimates relative position between a camera and dynamic objects by SLAM. • Dense optical flow is leveraged to densely reconstruct and track non-rigid objects. • The proposed method demonstrated superior performance compared to a deep learning-based monocular depth estimation.