基于Kinect的移动机器人视觉SLAM研究

发布时间：2019-06-14 16:01

【摘要】：智能移动机器人要求能够在复杂环境中自主导航与定位,而同时定位与建图(SLAM)是实现移动机器人完全自主移动的前提和关键。基于视觉的SLAM技术由于其价格低廉、信息丰富及特征易提取等优点受到了研究者们的广泛关注,由于Kinect相机能够方便、快速地获取环境的RGB-D信息,因而被大量的应用于视觉SLAM中。目前采用RGB-D类传感器的主流视觉SLAM方案由图像处理前端和位姿优化后端组成。针对视觉SLAM系统实时性问题,论文主要在前端图像处理部分,对影响系统实时性的关键性环节进行研究并提出改进方法。前端的图像处理效率会直接影响到整个SLAM系统的实时性,论文引入光流法来快速跟踪特征点在图像间的运动,并与传统的特征匹配法进行对比分析,提出了一种光流和特征匹配法组合的方法。其中,在运动估计部分,采用光流法来实时地估计移动机器人的运动;为了消除运动估计过程中的累积误差,采用基于特征匹配法的回环检测来增加机器人位姿间的约束。此外,为了提升后续的位姿优化效率,在回环检测过程中采用一种局部回环和随机回环组合的策略。在后端位姿优化部分,根据运动估计和回环检测部分得到的机器人位姿及位姿约束,采用了g2o算法对机器人位姿进行全局优化。论文通过实验对比分析了光流法和特征匹配法的性能,基于基准数据集的实验结果表明组合算法与传统的特征匹配法相比,在保证了SLAM系统定位精度的前提下,运行效率提高了28.5%,有效地提高了视觉SLAM系统的实时性。最后,实际场景在线实验结果表明,论文方法能够实时、准确地估计机器人的运动轨迹和构建室内场景的三维地图。
[Abstract]:Intelligent mobile robot requires autonomous navigation and positioning in a complex environment, while the simultaneous positioning and mapping (SLAM) is the premise and the key to realize the complete autonomous movement of the mobile robot. The visual-based SLAM technology is widely concerned by the researchers because of its low price, rich information and easy extraction of features. Because the Kinect camera can easily and quickly get the RGB-D information of the environment, it is widely used in the visual SLAM. At present, the mainstream visual SLAM scheme with RGB-D sensor is composed of image processing front end and pose optimization back end. In view of the real-time problem of the visual SLAM system, the paper mainly focuses on the front-end image processing part, and studies the key link of the real-time performance of the system and puts forward the improvement method. The image processing efficiency of the front end directly affects the real-time performance of the whole SLAM system, and the paper introduces the optical flow method to track the motion of the feature point in the image rapidly, and compared with the traditional feature matching method, and provides a combination method of the optical flow and the characteristic matching method. In the motion estimation part, the optical flow method is adopted to estimate the motion of the mobile robot in real time; in order to eliminate the accumulated error in the motion estimation process, a loop detection based on the feature matching method is adopted to increase the constraint of the position of the robot. In addition, in order to improve the subsequent pose optimization efficiency, a local loop and a random loop-back combination strategy is used in the loop detection process. In the rear-end pose optimization part, the robot pose and pose constraints are obtained according to the motion estimation and the loop detection part, and the global optimization of the robot pose is carried out by using the g2o algorithm. Based on the experimental results of the reference data set, the performance of the optical flow method and the characteristic matching method is analyzed and compared, and the operation efficiency is improved by 28.5% under the premise of ensuring the positioning accuracy of the SLAM system compared with the traditional feature matching method. And the real-time performance of the visual SLAM system is effectively improved. Finally, on-line experimental results of the actual scene show that the paper can estimate the robot's motion track and build the three-dimensional map of the indoor scene in real time.
【学位授予单位】：南昌大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】