Research On Three-dimensional SLAM Algorithm Using Kinect And Visual Dictionary

With the popularization and development of mobile robots, Robotics has been hailed as one of the 21st century top ten potential growth fields. However, as one of the key technologies in Robotics, SLAM (Simultaneous Localization and Mapping) faced great challenges. Since traditional filter methods, such as Kalman filter and Particle filter, would accumulate errors while solving SLAM problems. It could not be used in the process of large-scale map creation. With the deep study of SLAM, the emergence of graph based SLAM theory in recent years could solve the error accumulation. But graph based SLAM still faced many problem, such as sensors’pose, loop closure, graph framework, and so on. A three-dimensional SLAM algorithm based on BVW (Bag-Of-Visual Words) algorithm and Kinect which could improve the robust of the algorithm was proposed.Firstly, we analyzed the principles of common features, its disadvantages and advantages, and the selection of features in indoor scenes. SURF features which was accelerate by GPU was adopted. This approach could improve the speed of feature extraction and the robust of extreme scenes. Compared to traditional approaches based on RANSAC, a multi-strategy matching approach was proposed. Through cross-matching and kNN (k-Nearest Neighbor) matching algorithm to effectively reduce the initial match point, the approach improved the match speed.When comes to the camera pose calculation, common PnP (Perspective N Point) relied on the final matched points. ICP (Iterative Closet Point) converged slowly, and could easily fall into local minimum. An approach to calculate camera pose combined RANSAC SVD and G-ICP (Generalized ICP) was proposed. The simultaneous use of the image features and point cloud improved the robustness under extreme environments.Compared to commonly used algorithms like random selection and KD-Tree (k-Dimensional Tree), a tree structure based visual bag of words loop detection algorithm was designed which could greatly increase the speed of similar scene detection. At last, iSAM (incremental Smoothing and Mapping) Graph optimization algorithm was used to calculate the camera pose, the point cloud map and trajectory was created. It was shown that the algorithm can achieve good robust and precision.