Neural Network SLAM Algorithms for Dynamic Object Removal in Complex Environments

To address the issues of accuracy degradation or even operational failure of simultaneous localization and mapping (SLAM) algorithms in dynamic, complex, and low-texture environments, this paper proposes a target detection and dynamic point cloud removal algorithm based on a lightweight neural network. This algorithm integrates visual feature extraction with moving object detection to reduce the interference of dynamic targets in localization. A plane-matching technique is employed to overcome the lack of features in low-texture environments. Additionally, dynamic object 3D point clouds are removed using 3D data generated by laser scanning. On the KITTI dataset, the algorithm achieves a 53.59% reduction in root mean square error (RMSE) of absolute trajectory compared to the DynaSLAM algorithm, with significant improvement in dynamic point cloud removal accuracy. Results demonstrate the proposed algorithm's excellent localization accuracy, operational efficiency, and robustness.