A Visual SLAM Loop Closure Detection Method for Spraying Robots in Hilly Orchard

Abstract

In the challenging environment of hilly orchards, vegetation occlusion, and variable lighting conditions lead to sparse and unevenly distributed visual features. Moreover, the extensive area covered by these orchards results in a vast volume of images that require efficient real-time processing. This paper introduces an enhanced loop closure detection method tailored for hilly orchard settings to address these issues. The proposed approach integrates adaptive threshold-adjusted uniform ORB feature extraction with CLBP features through the CBAM attention mechanism, creating a novel feature descriptor that emphasizes key information. Furthermore, it employs HNSW to expedite feature matching and adopts MAGSAC++ to refine geometric consistency verification. Experimental evaluations on a dataset of 2,347 images from hilly orchards demonstrate that our method achieves a precision rate of 75.75% with an average runtime of 8.73 seconds. Compared to the traditional ORB vocabulary model, this represents a significant improvement: the precision rate has increased by 21.73%, while the runtime has been reduced by $\mathbf{1 6. 9 2 \%}$. These enhancements underscore the method’s superior real-time performance and accuracy in detecting loop closures within hilly orchard environments, thereby providing robust technical support for the efficient operation of spraying robots in such settings.