In order to improve the detection accuracy of cooperative target ball used for the precision assembly of large-scale devices by laser tracker in complex scenes, an efficient cooperative target ball detection method based on deep learning is researched. Firstly, the image features of the cooperative target are analyzed. Then, by using the improved YOLOv2 model, an improved method based on attention mechanism is proposed aiming to the cooperative target characteristics of multi-scale and large proportion of small targets. In order to improve the anti-interference ability of the network model in complex background, a method of data enhancement is also proposed. The test result shows that the proposed improved YOLOv2 network based on attention mechanism and data enhancement has strong anti-interference ability against complex background and significantly improves the detection accuracy of cooperative target ball. The detection accuracy on the cooperative target test set has reached 92.25%, which meets the target detection accuracy requirements of laser tracker in the large equipment precision assembly.