At present, with the development of the depth sensor, the on-line method of 3D reconstruction is realized. The difficulty of the online reconstruction is how to convert multiple dynamic overlapping depth maps into a three-dimensional model in real time under the premise of better reconstruction effect and reconstruction scale. To solve this problem, this paper constructed an efficient spatial voxel data structure based on the spatial index strategy, improving the hierarchical grid data structure of the traditional volume fusion method and realizing the real-time access to the hidden surface data. These surface data are only stored during the observation to achieve the sensor movement in the process of online scene reconstruction. The algorithm performs interactive real-time reconstruction of fine-grained detail and large-scale environment by completing several core stages, such as depth map preprocessing, camera pose estimation, depth map fusion and surface rendering, on the GPU (Graphics Processing Pipeline). The experimental results show that the average reconstruction time is 21.6 ms, the ICP pose estimation is 15 times, and the average frame rendering rate is increased by 13.3% when the reconstruction time is 8.0 ms and the reconstruction time is 21.6 ms, realizing the real - time 3D reconstruction of large scale scenes.