A Multi-Sensor Deep Fusion SLAM Algorithm Based on TSDF Map

Traditional 2D Simultaneous Localization and Mapping (SLAM) algorithms commonly use occupancy grid map models, which are susceptible to Gaussian noise. The constraint information in the backend optimization process is limited. Sensor data utilization at various stages is also incomplete. To address these issues, this paper proposes a multi-sensor deep fusion SLAM method based on the Truncated Signed Distance Function (TSDF) map. Firstly, the inertial sensing unit (IMU) is pre-integrated, and then the distortion correction of the laser point cloud is corrected by using the posture obtained after pre-integration. In the front-end, the Unscented Kalman Filter (UKF) method is used to fuse odometry, IMU data, and LiDAR scan matching results to obtain the pose information of the robot. The backend uses IMU pre-integration factor, loopback detection, and laser point cloud registration to enhance constraints for global map pose optimization and achieve deep fusion of multi-sensor data. The map model uses a TSDF map, which constructs obstacle edges through weighted fusion and linear interpolation, and it truncates the grid cells around obstacles, thereby reducing the influence of Gaussian noise. The performance of Karto-SLAM, Cartographer, and the proposed algorithm is verified by comparing the public dataset and the dataset collected in the real environment. The results show that the proposed method effectively avoids the ghosting phenomenon of traditionally occupied raster maps and reduces Gaussian noise in terms of mapping. In terms of positioning accuracy, the effect of back-end optimization is enhanced by a multi-constraint relationship, which reduces the relative and absolute pose errors of the real trajectory. Our method improves localization accuracy by an average of 9% compared to Cartographer and by an average of 34% compared to Karto-SLAM.