| Vision-based robotic grasping,utilizing machine vision technology for object recognition,localization,and manipulation,plays a significant role in various fields such as industrial production and home service,and has emerged as a frontier and hotspot in the field of intelligent robotics in recent years.Despite significant progress in vision-based robotic grasping technology,the presence of noise interference,visual occlusion,and hand-eye calibration errors,among other uncertain information factors in real working environments,makes precise and reliable grasping in complex application scenarios challenging.This thesis focuses on the problem of robotic grasping in uncertain information environments and proposes a grasping scheme based on image visual servoing,addressing three key issues: localization and tracking,feature extraction,and integration of software and hardware systems under uncertain information conditions.The main research contents are as follows:Firstly,regarding the issue of localization and tracking in uncertain noise environments and target motion states for robotic visual servoing,a control framework based on point feature image visual servoing was established.Furthermore,a localization and tracking algorithm integrating ProportionalDerivative Sliding Mode Control and Dual-Rate Extended Kalman Filtering was proposed,and its performance under different noise intensities and target motion states was evaluated through physical experiments.Secondly,building on the point feature image visual servoing framework to expand the application range and practicality of image visual servoing,a control framework based on line features was established.An Extreme Learning Machine was introduced to online tune PID control parameters,achieving stable tracking of static and dynamic targets.Next,addressing the problem of robust feature extraction in uncertain visual information environments,a keypoint detection method tailored for image visual servoing was proposed.Training samples containing various uncertain visual information scenarios were synthesized through domain randomization technology.A target detection and keypoint detection network were designed and cascaded to create a top-down two-stage keypoint detection network.Experimental analysis validated the effectiveness of the proposed method and revealed its good performance under various uncertain visual information patterns.Furthermore,based on the construction of the two-stage keypoint detection network,to further accelerate the extraction speed of target keypoint features,a single-stage multi-task learning keypoint detection network was proposed.This model utilizes a multi-task learning mechanism,constructing a backbone and neck network with shared parameters to simultaneously learn boundary box and keypoint information through a multi-task decoupling head.Quantitative and qualitative analysis confirmed the method’s high real-time and high accuracy performance in various uncertain information environments.Finally,by comprehensively applying the localization and tracking and feature detection modules introduced in previous chapters,a complete robotic image visual servoing hardware and software grasping system was constructed.Multi-scale evaluation metrics were designed,real uncertain visual information scenarios were created,and a series of physical grasping experiments were conducted to deeply analyze and verify the system’s comprehensive performance.This research provides theoretical references for robotic grasping under uncertain information conditions,offering new methods and ideas for robotic grasping technology,thereby further enhancing the grasping performance of robots in uncertain information environments. |
