A monocular pose estimation method based on semantic segmentation and rotating target detection is proposed to solve the problem of arbitrary pose, different size objects and discrete grasping angles. In the first stage, faster R-CNN(faster regions with convolutional neural networks features) is used for grasping detection to obtain candidate grasping boxes. Then, the semantic segmentation network is used to obtain the contour information of the object to be captured. Finally, the semantic segmentation results are used to select the most reliable grasping box for each object to be grasped, and the coarse angle estimation is completed at the same time. In the second stage, the fine grasping angle is obtained by rotating target detection to correct the deflection angle of the grasping frame in the first stage. In addition, considering the multi-scale characteristics of the grab, a multi-scale feature fusion module is proposed to make all layers of the pyramid share similar semantic features. Aiming at the problem of slow convergence speed in the later stage of iteration when the intelligent algorithm solves the inverse kinematics, the advantage of fast convergence speed of the Newton method is used to speed up the convergence speed of the intelligent algorithm in the later stage. Based on V-REP simulation and actual grasping detection experiments, the grasping detection accuracy of the proposed algorithm is 98.4%, and the actual grasping success rate is 88.3%. The harmful torque during simulation grasping is improved compared with that before correction, which can meet the grasping requirements of the manipulator.