Real-time Indoor 3D Reconstruction Based On RGB-D Image And Imu Data

The 3D reconstruction technology of indoor scene is one of the hot research fields in computer vision and robot.It is also an important part of autonomous navigation of mobile robot,reconstruction of unknown environment model and construction of semantic map.The 3D reconstruction technology of single vision sensor based on RGB-D camera is easy to generate high error position estimation under the condition of low texture scene and fast motion of sensor.With the time of reconstruction,the error of position estimation will accumulate gradually,resulting in a significant drift between the reconstruction model and the real environment.For a rapid motion in a short period of time,IMU can provide better position estimation results,and can effectively solve the above problems.Therefore,a visual-inertial 3D reconstruction system integrating RGB-D images and IMU data is designed in this paper.The main research contents are as follows:1.Design and calibration of visual-inertial 3D reconstruction system: this system visual sensor hardware ASUS Xtion Pro Live provides color images and depth information,and inertial sensor MPU6050 provides position data.First,the inertial parameters of RGB-D camera and the IMU are calibrated and error corrected respectively.Then,the Camera-IMU joint external parameter calibration method is studied and the validity of the method is verified by experiments.2.Research on multi-data fusion position estimation algorithm based on nonlinear optimization: For single-sensor 3D reconstruction technology based on RGB-D camera in the scene of fast sensor movement or lack of image texture characteristics(such as wall,ground,roof,etc.),based on the problem of increasing error of position estimation of visual odometer,the method of combining visual odometer and inertial odometer with nonlinear optimization algorithm is proposed to effectively improve the accuracy of position estimation,system robustness and 3D reconstruction accuracy of sensor.3.Research on map cumulative drift correction and loop detection method: A method of adding gravity vector to surface element model to improve the constraint and optimization of variation map is proposed,and the problem of map cumulative drift is solved.To solve the problem that the traditional random fern algorithm has poor robustness in global loop detection,a random fern algorithm based on map surface information is proposed.4.The effectiveness analysis of the visual-inertial 3D reconstruction system: the system is evaluated from the aspects of absolute trajectory square root error and reconstruction error in the dataset.At the same time,the robustness and feasibility of the visual inertial 3D reconstruction system are verified in the following four aspects:local and global loop detection experiments,map cumulative drift correction experiments,error analysis experiments under different motion conditions of sensors,and three-dimensional reconstruction real-time experiments.