To improve the efficiency and accuracy of automatic assembly of chip level atomic clock, a coarse-fine combined micro device pose detection method is proposed. Firstly, after analyzing the characteristics and application environment of the micro-devices, a rough localization method of micro devices based on SVM classifier trained by image gradient histogram features is proposed. At the same time, occlusion samples and fuzzy samples are constructed to solve the problem of missing detection in the case of defocusing and occlusion of micro devices, so as to improve the generalization ability of the model. Then, the LSD algorithm and linear growth method are used to extract the edge of the target and calculate the sub-pixel feature point coordinates of the image. Finally, the real-time position and posture of the micro device under the microscope field of view are calculated by the ICP algorithm. Experimental results show that the recognition accuracy of the core components of the device is 95.3%, the average positioning error is less than 26μm, the angle error is less than 0.5°, which can meet the needs of the robot to accurately grasp the targets in the micro assembly process.