Evaluation of RGB-D Multi-Camera Pose Estimation for 3D Reconstruction

Advances in visual sensor devices and computing power are revolutionising the interaction of robots with their environment. Cameras that capture depth information along a common colour image play significant role. These cheap, small, fairly precise. The provided, particularly point clouds, can be generated virtual environment, providing complete 3D for applications. However, off-the-shelf cameras often have limited field view, both on horizontal vertical axis. In larger environments, it is therefore necessary to combine from several or positions. To concatenate multiple clouds generate environment information, pose each camera must known outer scene, i.e., they reference coordinate system. achieve this, system defined, then every device positioned according this For cameras, calibration performed find its relation Several methods been proposed solve challenge, ranging structured objects such as chessboards features study, we investigate how three different estimation multi-camera perspectives perform when reconstructing scene 3D. We evaluate usage charuco cube, double-sided board, robot’s tool centre (TCP) position real case, where precision key define methodology identify points space measure root-mean-square error (RMSE) based Euclidean distance actual ground-truth point. reconstruction carried out using TCP produced best result, followed by cuboid; angled board exhibited worst performance.