This paper introduces a new approach of structure-from-motion (SFM) based on video sequence. This approach uses a series of new sensors to help gathering the information of camera pose. This idea significantly reduced the difficulty of 3D reconstruction while improving precision of the reconstruction. The other contribution of this paper is that we use GP-GPU to parallelize the computation of reconstruction process. We get significant performance improvement according to our test. The computer vision technologies used in this paper include but not limited to feature point tracker, multi-view geometry.The sensors include accelerometer and gyroscope, which could easily be found in many kinds of mobile devices. We will describe how to use these sensors to help recording the pose of camera.Feature tracking is a common application of computer vision, there are many kind of features can be extracted and tracked, for the consideration of efficiency and effort, we decided to use the classical KLT tracker in our application. We can compute the structure of the scene when we already known the poses of camera and the matched points among frames. This is achieved by following the multi-view geometry.The structure of GP-GPU is best fit for compute-intensive cases when the application can be parallelized. We use the OpenCL API to implemente the KLT tracker and triangulate method. |