AppStabLoc: YOLOv11s-FAST-LIO Fusion with DBSCAN-Kalman Optimization for Stable Apple Localization for Robotic Picking

Orchard harvesting robots face critical challenges in apple recognition and localization, including low efficiency, high costs, and inadequate anti-interference capability. To address these issues, this study proposes the AppStabLoc algorithm, inspired by human environmental perception. The algorithm integrates unmodified, vanilla YOLOv11s (real-time apple detection), FAST-LIO SLAM (spatial mapping and coordinate transformation), and DBSCAN-Kalman optimization (noise mitigation and precision enhancement). It employs YOLOv11s, which is combined with a proposed elliptical ROI to mitigate corner-related errors in rectangular ROI and quadratic weighted mean-based depth estimation to extract local 3D coordinates of apple centers, transforms these to global map coordinates via FAST-LIO to eliminate cumulative errors, and refines accuracy through DBSCAN-Kalman filtering for stable outputs. Experiments on a prototype equipped with LiDAR, IMU, RGB-D camera and edge computing unit demonstrated that optimized average RMSD reduced to 0.0009 m (87% reduction vs. raw data) and average maximum Euclidean distance to 0.0036 m (90% reduction), far exceeding the typical precision requirement ≤ 0.005 m for orchard harvesting robots. This low-cost, edge-computing-enabled algorithm overcomes limitations of standalone deep learning models and traditional approaches, providing a practical solution for high-precision stable apple localization in complex orchards and facilitating large-scale deployment of intelligent agricultural robots.