Abstract: To address the pose drift and mapping degradation caused by dynamic illumination and low-texture environments in farmland scenes for traditional visual SLAM(simultaneous localization and mapping), a stereo vision-based simultaneous localization and dense mapping method was proposed. First, line features were introduced into the tracking thread and combined with point features for fusion matching to enhance the system’s robustness in low-texture and dynamic illumination environments and improve the stability of feature extraction and tracking. Second, based on the original SLAM architecture, a dense mapping thread was added to generate high-precision disparity maps using a deep learning-based stereo matching network, effectively overcoming depth estimation errors in textureless, occluded, and edge regions. High-quality dense point cloud maps were constructed through point cloud registration, point cloud fusion, and point cloud filtering, and the map accuracy was further improved after global bundle adjustment (BA) optimization.Results showed that on the EuRoC and KITTI Odometry datasets, the average absolute trajectory error (ATE) of the StereoDenseSLAM (SDSLAM) algorithm was 0.1121 and 2.137, respectively, which were lower than those of ORB-SLAM2, ORB-SLAM3, and PL-SLAM algorithms, indicating a significant improvement in localization accuracy. On the self-built dataset, the SDSLAM algorithm achieved high-precision dense reconstruction results, which well reflected real farmland scene information and met the requirements for constructing 3D dense point cloud maps in farmland scenes.

Key words: stereo vision, farmland scenes, simultaneous localization, dense mapping, SLAM