With the continuous development of computer vision technology, people are paying more and more attention to the method of using computers to simulate actual 3D scenes, and the requirements for 3D reconstruction technology are getting higher and higher. Virtual and real fusion refers to combining the virtual environment generated by the computer with the actual scenes around the user through photoelectric display, sensors, computer graphics, multimedia, and other technologies. This is a technology that can obtain more convenient and direct expressions, and it is also a technique for expressing content more abundantly and accurately. The key to virtual and real fusion technology is the registration of virtual objects and real scenes. It means that the system should be able to correctly estimate the position and posture of the camera in the real world, and then place the virtual object where it should be. Machine learning is a multifield interdisciplinary subject that specializes in how computers simulate or realize human learning behaviors. It is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are in all the fields of artificial intelligence. This article introduces the virtual-real fusion 3D reconstruction method based on machine learning, compares the performance of the method with other algorithms through experiments, and draws the following conclusion: the algorithm in this study is the fastest, with an average speed of 72.9% under different times. To evaluate the image acquisition indicators of each algorithm, the algorithm in this study has the lowest error rate. The matching accuracy of each algorithm is tested, and it is found that the average matching accuracy of the algorithm in this study is about 0.87, which is the highest.