Real-Time Navigation Line Extraction for Cabbage Harvesting Robots in Complex Farmland Environments: A Multi-Mask Fusion and Dynamic Row-Tracking Approach

Abstract

As an important leafy vegetable economic crop in China, cabbage harvesting is faced with problems such as low efficiency of manual operation, high labor intensity and high labor cost. However, the existing navigation technologies for harvesting robots has shortcomings in accuracy and adaptability to complex environments.This research presents an advanced navigation line extraction framework tailored for automated cabbage harvesting, targeting the challenges posed by mature heading-stage crops in diverse field conditions. By introducing a multi-mask fusion segmentation strategy grounded in the HSV color model, the study overcomes the limitations of conventional grayscale methods, enabling precise separation of cabbage vegetation and plastic mulch while effectively suppressing interference from weeds, soil reflections, and plant yellowing. Three key algorithmic innovations are developed: First, an adaptive row-tracking algorithm incorporates dynamic historical midpoint adjustment and perspective-constrained hierarchical scanning, achieving an average angular accuracy of 0.24° for central navigation lines across high- and low-light environments, with a 94% reduction in false edge detections compared to traditional sliding window techniques. Second, a RANSAC-based robust fitting method reduces boundary line extraction time to 0.192 seconds per frame and centerline fitting time to less than 0.001 seconds, demonstrating superior performance in handling complex scenarios such as missing plants, curved rows, and heading angle deviations up to 15°. Third, a vector bisector-based centerline extraction approach, utilizing geometric intersection modeling and direction vector synthesis, ensures navigation lines maintain an accuracy within 2 pixels relative to manual annotations. Experimental validation in Xiangcheng County, Henan Province, using the ‘Zhonggan 56’ cabbage variety, confirms the framework's effectiveness, achieving an average precision rate of 92% (defined as the ratio of correctly detected rows to total annotated rows) across varied field conditions including weedy plots and camera tilts. These findings establish a reliable foundation for automated cabbage harvesting systems, with potential applications extending to other leafy vegetable crops. Future research will focus on field deployment via edge computing, followed by algorithm optimization through parallel architectures and deep learning integration and GIS technologies for large-scale agricultural management.