Outdoor high-precision 3D dense mapping system based on stereo visual SLAM

Traditional visual SLAM algorithms, which build maps based on sparse reconstruction, can hardly meet the demands for autonomous navigation and obstacle avoidance in outdoor environment. Therefore, we propose a visual SLAM construction algorithm to address this issue. The algorithm employs a deep stereo matching network based on a cross-attention mechanism and an adaptive disparity refinement strategy to suppress mismatched disparities, thereby enhancing the robustness of dense depth estimation in complex environments. Additionally, a separate dense mapping thread is applied to fuse the refined dense depth with SLAM poses for loop detection and pose correction, generating high-quality color depth point clouds and reconstructing high-precision 3D dense maps. To evaluate the performance of our algorithm, two experiments are conducted using the KITTI dataset and field test data, respectively. The experimental results show that our algorithm achieves a relative depth estimation error of 8.236% with the field test data, while 92.902% of the dense points in reconstruction 3D map have an error within 0.443 m. These results demonstrate that our 3D mapping algorithm achieves higher accuracy and robustness than traditional methods and solves the problem of sparse point clouds and lack of environmental information in traditional visual SLAM systems.