In the first part of this Master’s Thesis, literature survey is done to find the best performing algorithm in terms of real-time performance... Show moreIn the first part of this Master’s Thesis, literature survey is done to find the best performing algorithm in terms of real-time performance and quality of 3D reconstruction from depth images. Kinfu of PCL is found to be the best in terms of real-time performance and maintainability, as it is open-source software, with thousands of users working on to improve it. The primary part of the thesis is to improve on the 3D reconstruction quality of Kinfu and this is achieved by improving on the three important stages of Kinfu, Depth Filtering, Pose-estimation and 3D reconstruction. A new Anisotropic depth filtering kernel is proposed and it is found to give 23% quantitative improvement over baseline Kinfu. A modified pose-estimation which incorporates the uncertainty or error characteristics of the depth measured as part of the ICP algorithm is proposed. This results in a 51% quantitative improvement and a big subjective quality improvement over baseline Kinfu. A similar approach based on depth error characteristics but applied to 3D reconstruction stage which is based on an already published paper [1], is implemented. All the improvements in each of the three stages are combined together, to get a robust, real-time 3D reconstruction framework. M.S. in Electrical Engineering, May 2015 Show less
