The Real-time Camera Localization And Three-dimension Mapping Method Based On RGB-D Image Sequences

Simultaneous Localization and Mapping problem has been a hot topic in the field of multiple disciplines in recent years, along with civil UAVs, home service robots, as well as the UGV of Google, the holographic of Microsoft and other high-tech concept of products, making itself attract more and more attention by the major universities and enterprises. This one, because the visual simultaneous localization and mapping problem take the vision sensors as a tool to perceive the scene environment, especially compared to conventional sensors, if they were influenced by some unfavorable factors, the sensor could accurately locate the current position taken the globally cumulative errors into consideration. Therefore, our article has proposed and implemented a realtime method of simultaneous localization and mapping based on RGB-D visual information.In this paper, we use the RGB-D image sequences as a visual input, separate the simultaneous localization and mapping problem into two main tasks which can compute in parallel. One part is based on visual odometry design, the major function in this part is estimating the trajectory of the camera in realtime, we use the ORB feature detection and matching algorithm between image frames for feature tracking, ensuring the realtime feature extraction algorithm, and use PROSAC algorithm between image frames, add motion constraints to ensure that the trajectory calculated sufficiently accurate; the other part is reconstructing a three-dimensional map of the scene simultaneously, where we use a keyframe screening method to weed out redundant information in the reconstruction process, and use a probabilistic occupancy estimation based on octrees to represent the three-dimensional map of the scene, to improve the storage efficiency of our SLAM. Moreover, throughout the simultaneous localization and mapping process, we use the topology optimization of ways to reduce the accumulated error generated by the method during operation. Finally, our method performs effective, fast and accurate in SLAM problem by the results of experiments.