Research On Dense Visual Odometry SLAM Based On Kinect

Simultaneous Localization and Mapping(SLAM) problem is considered for a mobile robot to make autonomous navigation and obstacle avoidance. Currently solutions to the problem of SLAM are classified into based-on filter and based-on graph. In recent years based-on graph SLAM method is more popular, which can be divided into two parts: the front-end and the back-end. The front-end is mainly responsible for the environment construction and the back-end for the environmental map optimization. Dense Visual Odometry SLAM(DVO_SLAM) is a kind of based-on graph SLAM solutions. In the last few years the sparse visual odometry method is popular based on visual feature points. Compared the based-sparse SLAM solution, dense visual odometry SLAM makes full use of the image information. DVO_SLAM method is based on the assumption of the consistency principle of light, namely for the same point P in the world, the light intensity of the point P is the same in different view of the camera image.First, this thesis studies and analyzes DVO_SLAM method. We respectively analyze the complexity and real-time performance in principle and from experimental results. Then we find that it has high complexity and its real-time performance is not well. After analyzing the original DVO_SLAM method, this thesis find out the corresponding improvement parts, including the selection strategy of key frame, image prediction and back-end optimization. Then, the original algorithm is not well on accuracy of image prediction. This thesis proposes a more precise pixel interpolation method to improve the accuracy of image prediction. At the same time, the three strategies for selecting key frames in the original algorithm cannot be good balance between efficiency and performance relationship. This thesis proposes a new strategy based on effective pixels to select key frames. Finally, through experimental analysis, we find that the original DVO_SLAM method performs not well for some datasets containing error closed-loop constraints and so we adopts a dynamic covariance scaling algorithm to eliminate error closed-loop constraints in the back-end optimization.In the end, after a series of experiments on standard datasets, we prove that the improved dense visual odometry SLAM method performs better than other sparse visual characteristics SLAM methods on both efficiency and performance. The improvements on the original algorithm are also very good to improve the performance of the algorithm.