Weeds Automatic Annotation and Stem Localization Based on Spatial Association for Laser Weeding Robot

Abstract

In precision agriculture, the application of artificial intelligence and high-power laser technology for weed control offers significant efficiency and accuracy advantages. Current laser-based weed control systems encounter limitations in data quality, annotation efficiency, and the spatial precision of weed stem localization. In addressing these challenges, this study introduces an innovative framework for weed detection and automated annotation, alongside a localization algorithm for weed stems. The methodology enhances the accuracy and automation of weed annotation across sequential imagery through the optimization of the weed object detection algorithm and zero-shot segmentation techniques, coupled with spatial correlation analysis utilizing three-dimensional (3D) weed reconstruction data. For stem localization, regression predictions of weed and stem positions are obtained via a two stage segmentation network. Spatial association and constraint mechanisms are then applied based on motion cues and tracking data to correct the results. Experimental evaluations on the dataset demonstrate that the automated annotation accuracy reaches 89.2%, with an average pixel error within 5.3 pixels. The positional accuracy for weed stem localization attains 92.7% within the permissible pixel error margin, representing improvements of 5.34% and 1.2% over UniStemNet and EffiStemNet, respectively. These results validate the efficacy of the proposed algorithm and offer valuable insights for the advancement of laser-based robotic weed control systems in precision agriculture.