Research And Application Of Mobile Robot 3D Mapping And Navigation With RGB-D Perceptions

The application of Kinect sensors in the perception technology of mobile manipulative robot has become a hot research topic, in which a Kinect sensor can provide rich information of RGB and Depth data, namely RGB-D information. However, a Kinect sensor provides large, redundant and noisy data in low resolution. The problem of 3D environment mapping with higher precision and efficiency for reliable robot localization and navigation in its long-term operation is a key problem to be solved. Therefore, in this dissertation, the research and application of robot 3D mapping and collision avoidance navigation with RGB-D perceptions is studied, and a Kinect based robot indoor navigation system is also developed.Firstly, a process of RGB-D data pre-processing and environmental feature extraction is designed, which provides the basis for 3D mapping and navigation. The process deals with the problem of large, redundant and noisy data in raw Kinect point clouds. Secondly, an improved method of 3D mapping based on a spatial bag-of-word representation of scene images, namely SDBoW2 model, is put forward. To be more specific, the proposed SDBoW2 model represents scene features in binary vectors and incorporates spatial information of ORB descriptors. Thus by using the proposed SDBoW2 model, the process of loop-closure detection and 3D point cloud registration is implemented with higher accuracy and efficiency. Meanwhile, the memory required for the computation of the 3D global mapping is also reduced, which allows more reliable 3D mapping and localization in large indoor environments. Thirdly, a RGB-D based real-time navigation system is developed that contains two strategies for two different types of indoor environments, respectively, In the situations that a 3D global map is available, a first strategy is developed that employs the result of RGB-D based localization. The strategy uses the output points of global path planner as navigation reference points, and uses the Nearness Diagram (ND) algorithm based on RGB-D perceptions for obstacle avoidance in complex environments. In other situations that the environments are visually featureless (e.g., corridors) and a 3D dense map cannot be established using the proposed mapping method, an alternative strategy is developed. The strategy makes use of the grid and geometrical features of the environments, in which the navigation reference points are acquired by extracting the skirting lines from the RGB-D data, and the local motion control variables are also calculated by the Nearness Diagram algorithm. As an example, a wall-following and collision avoidance behavior of robots in corridors is achieved.Lastly, the application system of indoor mobile robot with RGB-D perceptions is established, in which basic functional modules of environmental feature extraction,3D mapping, vision localization and collision avoidance navigation are developed. The feasibility and effectiveness of this system are verified by experimental results.