Camera pose estimation for 3D scene reconstruction based key frame extraction

Estimation of camera pose is a fundamental problem in many computer vision areas, such as simultaneous localization and mapping, robot navigation, 3D scene reconstruction, etc. Recently, using low-cost RGB-D camera to perform 3D scene reconstruction is an active area of mobile robotics research. The ability to localize a camera moving is a very important step. To estimate the camera trajectory, we need to compute the geometry relationship between a set of images. Regardless of which 3D reconstruction algorithm is used for an application, the quality of the results depends heavily on the quality and consistency of the input image sequence, which are required to approximately cover the entire scene and at the same time minimize the content redundancy between the selected frames. Therefore, we introduce the use of a key-frame selection strategy as a preprocessing technique, which not only greatly saves the computation time, but also helps significantly reduce the number of repeated features to improve the camera pose estimation quality. The key frame selection strategy that has been used in this research utilizes the pixel intensity differences among subsequent frames to automatically select only the frames that contain the desired contextual information and discard the rest of the insignificant frames. Our pose estimation for 3D reconstruction system has been applied successfully to video from handheld RGB-D camera and a RGB-D camera mounted on a ground robot. The performance of the proposed technique is observed to be significantly improved using our key frame selection strategy.https://ecommons.udayton.edu/stander_posters/1946/thumbnail.jp