Visual-LIDAR SLAM Based on Supervised Hierarchical Deep Neural Networks

Deep learning has become increasingly crucial in simultaneous localization and mapping (SLAM). Supervised deep learning SLAM methods need ground truth data for training but achieve highly accurate predictions. Such training techniques are used by some visual or LiDAR SLAM methods. However, these methods mostly rely on single-sensor data, missing the combined advantages of LiDAR and visual information. Multisensor fusion SLAM merges diverse data types, enriching features and boosting performance. This paper presents a new deep visual-LiDAR SLAM method that integrates both visual and LiDAR data. The SLAM system includes a deep visual-LiDAR odometry module, a deep learning-based loop closure detection module, and a 3D mapping module. The deep visual-LiDAR odometry module uses a hierarchical feature encoding module to capture features from two frames of color point clouds at various levels. An attention feature decoding module employs an attention mechanism to merge these features, determining their relative pose. The iterative pose optimization module continuously refines the pose accuracy, improving robustness to outliers. The loop closure detection module uses a deep learning-based global descriptor for precise positional matching. The 3D mapping module constructs the environmental map. Experimental results on the KITTI odometry dataset show that our method outperforms other supervised and traditional geometric methods, with lower rotation and translation errors. By integrating visual and LiDAR data, our SLAM system achieves higher accuracy and robustness in pose estimation compared to single-modal SLAM systems.