A visual-inertial localization algorithm for greenhouse robots based on point-line features

Abstract

Dynamic illumination conditions and potential obstacles in greenhouse environments can severely compromise the localization accuracy of point feature-based visual-inertial systems. To address these challenges, this study proposes a point-line feature-based visual-inertial localization algorithm that integrates an improved EDLines detector with an adaptive adjustment strategy. By embedding the enhanced EDLines line detection module into the front-end processing stage of the framework, the algorithm strengthens feature extraction capabilities, effectively mitigating the limitations imposed by sparse point features and suboptimal feature matching commonly encountered in greenhouse environments. To further reduce the adverse effects of low-quality line feature matches—particularly those arising from short line segments—this study introduces a length-based filtering mechanism combined with an optimized matching strategy, thereby significantly improving localization accuracy. To address fluctuations in feature quantity caused by varying illumination conditions, an adaptive strategy is proposed to dynamically adjust the number of extracted point and line features, thus maintaining localization robustness. A custom dataset was collected in a real greenhouse environment, covering both sufficient illumination and inadequate illumination conditions, and integrated with simulation data for comprehensive algorithm validation. Simulation results demonstrate that the proposed method achieves a maximum localization error of 0.152 m, surpassing the baseline algorithm in terms of accuracy and confirming its suitability for greenhouse localization tasks. Additionally, experimental evaluations conducted on the self-collected dataset reveal that, under sufficient illumination conditions, the average errors along the X and Y axes are 0.054 m and 0.114 m, respectively. In inadequate illumination environments, the proposed algorithm reduces the average X-axis and Y-axis errors by 80.0% and 82.5%, respectively, compared to the baseline method, demonstrating a substantial improvement in localization performance. Furthermore, the overall tracking and matching time remains below 40 ms, fulfilling the real-time operational requirements for greenhouse applications.