DeepVIO:3D幾何制約を用いた単眼視覚慣性オドメトリの自己教師あり深層学習
DeepVIO: Self-supervised Deep Learning of Monocular Visual Inertial Odometry using 3D Geometric Constraints
ステレオ画像から得た3Dオプティカルフローと6自由度姿勢を教師信号とし、2DオプティカルフローとIMUデータを融合する自己教師あり単眼視覚慣性オドメトリを提案。KITTIとEuRoCで高精度かつキャリブレーション誤差やデータ欠損に頑健であることを示した。
著者: Liming Han, Yimin Lin, Guoguang Du, Shiguo Lian
分類: cs.RO, cs.CV
原文アブストラクト
This paper presents an self-supervised deep learning network for monocular visual inertial odometry (named DeepVIO). DeepVIO provides absolute trajectory estimation by directly merging 2D optical flow feature (OFF) and Inertial Measurement Unit (IMU) data. Specifically, it firstly estimates the depth and dense 3D point cloud of each scene by using stereo sequences, and then obtains 3D geometric constraints including 3D optical flow and 6-DoF pose as supervisory signals. Note that such 3D optical flow shows robustness and accuracy to dynamic objects and textureless environments. In DeepVIO training, 2D optical flow network is constrained by the projection of its corresponding 3D optical flow, and LSTM-style IMU preintegration network and the fusion network are learned by minimizing the loss functions from ego-motion constraints. Furthermore, we employ an IMU status update scheme to improve IMU pose estimation through updating the additional gyroscope and accelerometer bias. The experimental results on KITTI and EuRoC datasets show that DeepVIO outperforms state-of-the-art learning based methods in terms of accuracy and data adaptability. Compared to the traditional methods, DeepVIO reduces the impacts of inaccurate Camera-IMU calibrations, unsynchronized and missing data.
