Visual Inertial Odometry Based on Deep Features

Visual odometry is the cornerstone in the field of SLAM. Monocular visual odometry occupies an important position because of its low cost and less camera calibration, but it has some shortcomings such as scale uncertainty, scale drift, poor robustness, and so on. To solve these problems, based on ORB-SLAM3, we process a monocular visual-inertial navigation odometer with depth features, referred to as DF-VIO (Visual Inertial Odometry Based on Deep Features) , which uses depth features extracted by deep learning network to replace traditional artificial point features, and fuses artificial line features to enhance the robustness of the system in real complex scenes. Besides, the system provides a variety of pose tracking methods. In addition to the method based on the constant speed model and tracking reference keyframe, a pose tracking method based on the predicted repeatability map is also provided, which further improves the accuracy of system pose tracking. Comparative experiments are carried out on the open data set EuRoC, and the average trajectory error is reduced by 25.9% in pure visual mode and 8.6% in visual-inertial mode, which proves that the system proposed in this paper can be more robust in complex scenes.