【作者】 冯肖维；

【导师】 方明伦； 何永义；

【作者基本信息】 上海大学 ， 机械制造及其自动化， 2011， 博士

【摘要】 随着计算机科学、传感器技术、人工智能等学科的发展和机械设计制造水平的不断提高,移动机器人日益向着智能化和自主化的方向发展。机器人实现自主地行走,执行具体的任务,必须具有判断自身位姿的能力。位姿计算或定位问题一直是移动机器人领域的研究热点,受到国内外的广泛关注。本论文针对现有移动机器人位姿计算方法存在的问题展开研究,重点对里程计、激光和视觉传感器的信息处理与融合、固定环境中多机器人位姿计算、移动机器人自主位姿计算等内容进行了深入的研究。主要内容包括以下几方面:针对机器人位姿计算中,激光测距仪所获得的原始距离图像在景物空间呈现多尺度特性,使得特征提取过程容易出现虚假特征和特征丢失问题。本文提出基于特征估计的多尺度自适应滤波方法,对距离图像进行滤波处理,并根据图像的局部曲率对特征进行分割与辨识,有效减少了虚假特征和特征丢失情况的发生。实验表明,该方法能够提高二维激光距离图像特征提取的成功率,从而增加机器人位姿计算的精度和鲁棒性。针对视觉传感器原始采集图像存在畸变和干扰特征,给动态环境中的机器人辨识带来困难的问题。本文首先建立包含径向和切向畸变的广角镜头成像模型,对失真图像进行校正;然后利用颜色分割和形状模板匹配相结合的特征识别算法,提高了机器人辨识的可靠性。同时依靠机器人运动模型预测其在下一时刻的位置,从而减少图像的搜索范围,为基于多摄像机的机器人位姿计算提供了实时性保证。针对多传感器信息融合能够提高机器人位姿计算的精度和可靠性问题,本文提出了分布式多传感器信息融合位姿计算方法。通过数据关联技术将传感器测量数据与环境中机器人相匹配,利用分布在各机器人客户端上的双层无嗅卡尔曼滤波器将来自视觉传感器和激光测距传感器的信息与来自码盘的信息相融合,解决了固定环境中多机器人位姿计算问题。该分布式多传感器融合位姿计算方法不受机器人数量的限制,扩展性强。实验表明,本文所述方法具有较高位姿计算精度和较强的稳定性。针对现有基于粒子滤波的机器人自主位姿计算方法存在算法效率低、粒子退化等问题,本文利用最小偏度采样无嗅卡尔曼滤波(Minimal Skew UnscentedKalman Filter,MS-UKF)将最新的传感器观测数据融入粒子滤波的采样函数中,使粒子滤波融合算法即使在减少粒子数量的情况下,依然保持较高的计算精度。并在粒子滤波重采样过程中将MS-UKF作为辅助,有效地降低了粒子退化现象。最后在移动机器人上进行了实验,结果表明,本文所述改进算法能够有效提高位姿计算精度和效率。本论文有关移动机器人位姿计算的问题研究,将有助于智能移动机器人环境感知、协同控制、自主导航等能力的提高,这将对拓展移动机器人的应用领域,具有积极的理论和实际意义。更多还原

【Abstract】 With the development of computer science, sensor technology, artificial intelligence and the improvement of manufacture level, the robotics increasingly tends toward intelligent and autonomous. In order to make the robot move in the environment autonomously and do the task, the robot must be capable of calculating its pose. Pose calculation or localization problem is a key researching domain in the mobile robot community and get much attention around the world.This dissertation is focused on the multi-sensor fusion based pose calculation problem for mobile robot. The intention of this dissertation is to describe the research on encoder, laser rangefinder and vision data processing and fusion, multi-robot pose calculating and tracking problem in fixed environment and the mapping based localization problem for autonomous robot. The main contents and contributions of this dissertation include the following aspects:In pose calculation of mobile robot, the original range image from Laser rangefinder appears at non-uniform scale or resolution in scenery, which causes false alarms and missed detections. An adaptive smoothing algorithm within a scale space framework is introduced for noisy range image of laser rangefinder in order to extract features. Then the features can be segmented and identified according to the curvature of the range data, which decrease the false alarms and missed detections. Experimental results show that the proposed method is efficient in feature extraction,which can improve the accuracy and robustness of robot pose calculation. When mobile robot working in dynamic environment, the original vision image has the disadvantage of distortion and contains disturbing features, which lead to difficulty for robot pose calculation. In order to solve the problem, a flexible camera model contained radial and tangential distortion is established to correct the distorted images. Then this paper uses a recognition algorithm combined color segmentation with recognition method based on a shape template, which effectively reduce misidentification and improve the robustness of robot recognition. Then, a prediction algorithm based on the model of mobile robot is presented. This method can predict pose state of robot in the next frame and reduce the searching area of image, which guarantee the real-time performance for the pose calculation.Multi-sensor fusion can improve the accuracy as well as the robustness of the pose calculation for mobile robot. In order to calculate the poses of several robots, a distributed multi-sensor fusion pose calculation method is proposed. The measured data from the vision system and laser rangefinder is matched and correlated with the robots in the environment by data association process, and are combined with the information from encoder by a two layer UKF on robot. The distributed framework takes the advantage of high flexibility, and does not limit to the number of tracking robots. Experimental results show that the proposed method has high accuracy of robot pose measuring and strong stability.In order to improve the precision and reduce the sample impoverishment problem of autonomous localization based on Particle Filter, an improved Rao-Blackwellized Particle Filter by incorporating the most recent sensor observation is proposed. The filter uses Minimal Skew Unscented Kalman Filter (MS-UKF) to generate proposal distributions in order to optimize the samples, which can obtain satisfying calculation results with a small sample set. Moreover, we propose an MS-UKF based assistant-proposal distribution during resampling, which keeps the diversity and randomness. A series of experiments are carried out on a mobile robot, and the results show that the method effectively improves the precision and efficiency of robot pose calculation.The contribution of the dissertation consists in improving the capability of environment perception, co-operating and autonomous navigation of mobile robot. This is of positive academic significance and practical importance to improve the quality and wide the application field of mobile robot.更多还原