Dense Localization of a Monocular Camera Using Keyframes

In this paper, we present a low cost localization system that exploits dense image information to continuously track the position of a camera in 6DOF. It leverages of the use of a set of selected "key frames" separated in distance from which a depth map is available to create a local 3D point cloud. In this way, we avoid the computational overload caused by common dense sequential approaches. The system uses a 3D-2D technique to calculate an initial pose estimate for the intermediate camera frames. A refinement step stated as a Non Linear Least Squares (NLQs) optimisation is performed by minimising the photo-consistency error. The NLQs cost function is defined by aligning a warped image and an image associated to the closest key frame. The minimum solution is calculated using the Levenberg-Marquardt method. To validate the accuracy of our system, we conducted experiments using data with perfect ground truth. Our assessment shows that our system is able to achieve up to millimeter accuracy. Most of the expensive calculations are carried out by exploiting parallel computing and GPGPU.