An Improved 3D Reconstruction Algorithm Based On The Inertial Sensor

This thesis researches indoor scene reconstruction based on the depth camera.The inertial sensor is applied to 3D reconstruction to improve the stability of camera tracking,reduce the pose accumulation error and running time of reconstruction.First,this thesis researches the problem of tracking failure.In the original 3D reconstruction algorithm,the camera pose is calculated by the ICP(Iterative Closest Point)algorithm.When the depth camera shake occurs during the reconstruction process,ICP algorithm will fall into the local minimum value,leading to tracking failure.In order to solve this problem,the pose obtained from the inertial sensor is taken as the initial value of the ICP algorithm to prevent ICP from falling into local minimum and improve the robustness of the algorithm.Then,this thesis researches the problem of pose accumulation error.Due to the presence of the sensor noise,the error exists in the point cloud data after the conversion.Besides,if the feature information in the space is not enough,the point cloud matching will not get a perfect match,which makes the error exists in the matching result.When scanning for a long time,the error will accumulate,resulting in failure of the reconstruction process.In order to solve this problem,the pose obtained by the inertial sensor and the pose calculated by ICP are fused by Invariant Extended Kalman Filter(Invariant EKF)to obtain a more accurate pose.Finally,this thesis researches the real-time performance of 3D reconstruction.In the rendering step,raycasting is used to extract explicit surface data for visualization.The extracted surface data is also used to calculate the camera pose by registration with the point cloud of next frame.Thus,the pose difference between the current frame and the previous is calculated,as well as the pose difference between the current frame and the most recent raycasting.Then,by comparing with pre-determined thresholds,this thesis ignores the raycasting selectively to reduce the running time of reconstruction to guarantee the real-time performance,under the premise that the reconstruction accuracy is not affected.Experiments show that the improved algorithm tracks the camera accurately when the camera shake occurs,obtains a more accurate pose and improves the speed of the system.