The automatic control of excavator operation trajectories is a pivotal technology for autonomous excavators, with the essential prerequisite being the real-time measurement of manipulator poses. Given the complexity of the operating environment, traditional sensor-based measurement methods face limitations, whereas visual measurement emerges as a promising technique. Accurately measuring excavator manipulator poses involves a crucial aspect: mapping the relationship between image information and poses. First, to address the significant errors in pose prediction encountered with machine learning techniques like artificial neural networks, this work introduces a mathematical model for mapping this relationship, referred to as the pose mapping mathematical model, which includes calibrating model parameters. Second, to address the sensitivity of initial values in the calibration process, we propose a residual-guided initialization algorithm. This algorithm aims to ensure that initial values closely approximate the ground truth values, thus preventing matrix singularity at the source and avoiding parameter estimation divergence. Third, to tackle challenges such as unstable lighting conditions and discrepancies between the dataset and the mathematical model, we introduce the random sample consensus-driven Levenberg–Marquardt parameter optimization algorithm to enhance parameter estimation accuracy. Experiments with static and dynamic online measurement demonstrate that our method reduces pose measurement errors compared to existing methods. This research lays a solid foundation for developing visual measurement techniques for excavators and automated manipulator control based on visual measurements, also serving as a valuable reference for research on mechanical arms.
Read full abstract