Aiming at problems such as low recognition accuracy or poor localization ability of circular cooperative objects in low illumination or complex backgrounds in vision measurement, a model based on CNNs is proposed in this paper to optimize the YOLOv8 algorithm. The model designed in this paper has a total of 225 layers of network, about 3 million parameters and 8.2G FLOPs of computing power. The model is trained by using the circular cooperative target data set under different conditions, and the performance index and computational efficiency of the model are monitored in real time during the training process, and the model is adjusted and optimized in detail. The experimental results show that the algorithm has a precision of 99%, a recall rate of 92% and an average accuracy of 92%. Compared with traditional feature extraction methods such as Hough transform and YOLOv3, the accuracy of the proposed algorithm is improved by 14% and 4%. Recall rates increase by 17% and 2%. The average accuracy is improved by 10% and 2%. The algorithm can significantly improve the recognition and positioning accuracy of circular cooperative targets under variable conditions such as low illumination environment, complex background or small change of target shape.