基于三维视觉的机器人自主抓取技术是仓储物流自动化中的重要技术，在感知方面和控制方面都有较高的要求。在感知方面，要求三维测量系统可以精准、完整地测量待抓取物体的三维形貌，并快速准确地在作业场景中识别可行的抓取位姿；在控制方面，要求可以控制机器人执行器准确移动至指定的位姿完成抓取。针对基于三维视觉的机器人自主抓取技术在感知和控制方面的技术要求，现有研究还存在一定的局限性，因此本文以多视角面结构光测量系统作为三维测量系统，在感知与控制方面进行理论和技术研究。首先，面结构光测量系统的标定精度是测量精度的重要影响因素，而现有标定方法没有考虑镜头散焦和相机成像模型对标定精度的影响。本文从两方面提高标志点的定位精度，其一是通过对镜头散焦机理的分析，提出了高精度边缘定位方法；其二是使用相机成像模型，提出了模型驱动的标定点迭代补偿定位方法；从而提高了面结构光测量系统的标定精度与测量精度。其次，为提高复杂光学特性物体测量的完整性，本文设计了基于面结构光解码有效性掩模的解码测量与深度学习预测的融合测量方法，并针对现有基于三维成本体深度学习的多视角面结构光测量存在的分辨率低、精度差、且未考虑可见性的问题，提出了基于点云深度学习的可见性感知的多视角面结构光三维重建网络结构。再次，针对抓取位姿识别，本文研究了基于点云深度学习的单阶段抓取位姿识别方法，使用点云神经网络直接对测量点云进行处理并提取多尺度几何特征，同时预测各点的抓取质量和最优抓取位姿，提高了计算效率和对测量噪音的鲁棒性。最后，为了实现机器人执行器的准确控制，提出了基于面结构光相位图的视觉伺服控制方法，由相位图区域积分构建视觉特征提高了控制鲁棒性，使得可以用于无纹理物体并不依赖于环境光照，同时与使用三维点云进行控制相比降低了计算量，提高了控制效率。

3D-vision-based robot autonomous grasping technology is significant in warehousing logistics automation, which has high requirements in perception and control. In terms of perception, the 3D measurement system is required to accurately and completely measure the 3D shape of the objects to be grasped, and to quickly and accurately detect the feasible grasp pose in the work scene; in terms of control, it is required to control the robot end-effector move to the specified position accurately to grasp the target object successfully. Aiming at the technical requirements of perception and control of robot autonomous grasping, existing research still has certain limitations. Therefore, this thesis studies the theory and technology of perception and control by using a multi-view structured light measurement system as the 3D measurement system.First, calibration precision of structured light measurement systems is an important factor of the measurement accuracy. Existing calibration methods do not take into account lens defocusing and camera imaging model, which leads to deteriorated calibration precision. This thesis improves the positioning accuracy of the imaged reference points from two aspects: one is to propose a high-precision edge positioning method through the analysis of the lens defocusing mechanism; the other is to propose a model-driven iterative compensation positioning method for reference points by using camera imaging model with lens distortion. The proposed method improves the calibration precision and measurement precision of the structured light measurement system.Secondly, in order to improve the measurement completeness of objects with complex optical characteristics under different illuminations, this thesis designs a fusion measurement method that fuses the structured light decoding measurement result and deep learning depth prediction based on the structured light decoding validity mask. Existing cost-volume-based multi-view structured light depth prediction methods have the draw-backs of low resolution, poor accuracy, and no consideration of visibility. Therefore, this thesis proposes a point-based multi-view structured light depth prediction network structure with visibility-awareness.Thirdly, for grasp pose detection, this thesis studies a single-stage grasp pose detection method based on point cloud deep learning, which adopts a point cloud neural network to directly process the measured point cloud and extract multi-scale geometric features to predict the graspability and optimal grasp pose of each point simultaneously, improving the computation efficiency and robustness to measurement noise.Finally, in order to achieve accurate control of the robot end-effector, a visual servoing method based on structured light phase map is proposed, which constructs visual features from phase map area integration to improve the control robustness, so that it can be used for texture-less objects and does not depend on ambient lighting. Compared with control based on 3D point clouds, the proposed method reduces the computation cost and improves the control efficiency.