Dense Mapping Method for Grape Orchard Robots Integrating Semantic Segmentation

Abstract:

Aiming to address low localization accuracy, unreliable fruit recognition, and poor map quality in vineyard robots, PDS-SLAM, a dense mapping algorithm that integrated semantic segmentation was proposed. Built on ORB-SLAM3, each image was partitioned during feature extraction;the regional FAST threshold was adaptively adjusted according to regional corner counts;and quadtree uniformization method with minimum distance was applied, which improved spatial uniformity and matching robustness of feature points, thereby enhancing localization accuracy. A network, PDSNet, was proposed by integrating a DSA module into PIDNet, enhancing spatial perception of grape clusters and improving fruit recognition. A dense mapping thread and an octree thread were introduced: images were projected to recover local dense point clouds via a point cloud recovery algorithm;statistical outlier filter and radius filter were applied to remove aberrant points;semantic masks were used to annotate grape clusters, yielding a dense semantic map that was finally converted into an octomap. In experiments on the EuRoC dataset and a self-collected dataset, a 27.3% reduction in absolute trajectory error (ATE) on the MH03 sequence relative to ORB-SLAM3 and a 15.5% average increase in matched ORB features were achieved, indicating improved localization accuracy. PDSNet achieved an IoU of 78.9% for grape segmentation at 126.92 f/s. The results demonstrated that PDS-SLAM enhanced localization perception and produced dense semantic maps and octree maps, supporting autonomous navigation and precision operations for orchard robots.