Indoor 3D Reconstruction Based On RGBD Sensor

3D point cloud reconstruction is a key research topic in reverse engineering, how to build a precise, rich 3D scene model is an important research area. There are many methods of constructing a three-dimensional scene, usually we use laser range finder or depth camera to get the depth information, but the cost is relatively high and the standard laser range finder can only provide 2D cross-sectional, and can’t provide rich texture information. Recently, methods to deal with data of depth sensor are popular, this mainly due to the cheapness and good performance of RGB-D camera. For example the Kinect sensor which was launched by Primesense and Microsoft had changed this situation, making the 3D camera in an acceptable price.Firstly, we analyzed the traditional three-dimensional reconstruction method which based on ICP, we found the efficiency of direct process point cloud data was low, but in this paper, we use RGB-D sensor, obviously the 3D reconstruction method based on ICP has ignored important RGB information. So, we construct a 3D scene by tracking visual features, such as SIFT/SURF. And we use RANSAC to optimize the geometric transformation matrix which was computed by SVD, after this we use the matrix to reconstruct the scene. On this basis, we improved the visual feature extraction SURF algorithm, proposed Local-SURF. The Local-SURF algorithm is implemented by posture sensor and milemeter. We also compared the efficiency of SURF and Local-SURF, and we found Local-SURF was faster than SURF. The transformation matrix which computed from Local-SURF based 3D reconstruction algorithm can also used to correct the camera’s posture.According to the work in this paper, users reconstruct the indoor environment without expensive equipment, what they need are just Kinect and PC. As the result shown, although the 3D model is not accurate enough because of lower accuracy, the general outline is clear enough to meet the need in daily application.