In the proposed method, GPU-based SIFT features is used to realize real-time detection and matching between consecutive frames, their corresponding 3D points are optimized by comparing and fusing IMU data to improve odometry estimation stability, then keyframes are selected for incremental mapping, the graph optimization and moving least squares algorithm are applied to the dense reconstruction. To achieve a full 3D model of indoor environment, we realize the registration for over-lapped models by combining the features of key-frames and Simplified-ICP algorithm.
We plan to use the improved 3D SIFT feature points detection method, which combine with 3D RoPs feature descriptor to realize the estimation for the extrinsic parameters of over-lapped models, the K-Nearest Neighbor and Random Sample Consensus algorithms are used to optimize the accuracy of matched 3D feature points. We will make our code publicly available soon.
paper video code(coming soon...)