A real-time RGB-D registration and mapping approach by heuristically switching between photometric and geometric information

Real-time 3D mapping has many applications and has recently received a large interest due to availability of consumer depth cameras at low prices. In this paper, we present a 3D registration and mapping method that heuristically switches between photometric and geometric features, therefore allowing it to accurately register scenes that may lack either visual or geometric information. We propose a novel informative sampling based geometric 3D feature extraction technique in which the points carrying the most useful geometric information are used for registration. This increases the computational speed significantly while preserving the accuracy of the registration when compared to using the dense point cloud for registration. After extracting the features from sequential frames, they are assigned with feature descriptors and matched with their correspondences from the previous frame. The matches are then refined and a rigid transformation between the frames is calculated using a highly robust estimator. A global pose for the camera and associated 3D position of the points are computed by concatenating the estimated relative transformations and a 3D map is constructed. We evaluate our method on publicly available RGB-D benchmark datasets [1] and compare it to the state of the art algorithms.