【作者】 杨杰；

【导师】 张铭钧；

【作者基本信息】 哈尔滨工程大学 ， 机械电子工程， 2008， 博士

【摘要】 本文对视觉导航系统中视觉图像分割、视觉目标颜色特征提取、视觉目标定位方法等关键视觉技术进行了研究,并在此基础上建立了以路径边界引导的非结构化环境下的移动机器人视觉导航系统。在视觉图像分割研究方面,提出了基于子区生长的移动机器人视觉图像分割方法。该方法通过具有区域一致性的子区生长将自播种过程与图像分割结合在一起,代替传统自播种区域生长法中先确定种子点,后进行图像分割的过程,提高了分割效率。区域一致性判断对子区生长的应用至关重要,方差是常用的区域一致性标准,本文研究发现方差对图像噪声比较敏感,在理论分析其原因的基础上,本文定义了区域像素平均距离并提出以之作为区域一致性标准。理论分析和实验结果验证了区域像素平均距离比方差具有更好的噪声抑制能力和辨别区域一致性的能力。在目标颜色特征提取研究方面,提出了基于HSI颜色模型分量自适应量化组合的颜色特征提取方法。该方法利用HSI颜色模型对颜色描述的稳定性,结合基于直方图多阈值分类量化的自适应性,提取图像的颜色特征,实验结果验证了该方法的有效性。在直方图多阈值分类方法中,针对颜色分量直方图数据局部极大值引起的颜色分量过分类问题,提出一种去除颜色分量直方图局部极大值的方法。该方法定量分析颜色分量直方图数据的分布特点,从而确定并修改颜色分量直方图数据中的局部极大值。实验结果验证了该方法去除直方图局部极大值的有效性。基于平面单应矩阵估计的视觉定位方法的定位精度与目标的位置有关,应用于对移动机器人运动的平面上的目标定位时,对应每个摄像机的俯仰角度,必须建立相应的平面单应矩阵。针对这些问题,提出一种基于RBF平面单应估计的视觉定位方法,该方法将摄像机俯仰角作为RBF网络的输入之一,利用RBF网络的非线性逼近拟合图像平面到空间平面的非线性转换关系,通过RBF平面单应估计定位空间平面上的视觉目标。摄像机不同角度下的定位实验结果表明该方法能够对空间平面上的目标有效定位。对比定位实验表明,在相同条件下,该方法对空间平面上视觉目标的定位精度高于基于平面单应矩阵估计的方法。在图像分割、特征提取、以及视觉目标定位等关键视觉技术研究的基础上,提出了一种以路径边界引导的视觉导航方式。这种导航方式利用视觉处理技术获得移动机器人的路径边界信息,并从中提取移动机器人导航和避障控制参数。这种导航方式在引导移动机器人实现对非结构化环境的自主探索方面显示出较好的效果。更多还原

【Abstract】 In this dissertation, the visual image segmentation methods, color features extraction methods, and visual target locating methods for visual navigation system are researched. On the bases of those researches, a visual navigation system guided by the path boundary is established in non-structure environments.For visual image segmentation, a method based on sub-region growing for segmentation of visual image is presented. In this method, seeded process and image segmentation are combined by the growing of sub-regions which are coherent, instead of making certain seeded pixels before segmentation in traditional automatic seeded region growing method, and improves the efficiency of segmentation. Estimation of coherence of region is important for sub-region growing. Root-mean square error is usually used as standard of coherence of region, but its sensitiveness for image noise is discovered in our research. Based on the analysis for the reason of sensitiveness, the mean distance of region pixels is defined, and is used as standard of coherence of region. Theoretic analysis and experiment results show that mean distance of region pixels is better than root-mean square error on the capability of noise suppressed and region coherence distinguished.On color feature extraction of objects, a color features extraction method based on adaptive quantification and combination of components of HSI color model is presented. This method makes use of the description stability of HSI color model for colors, combining with adaptiveness of quantification based on multithresholding classification of histogram, and extracts color features of image. Experiments show that the presented method for color feature extraction is effective. In multithresholding classification of histogram method, aimed at the excessive classification problem of color components caused by local maximums of color components histogram, a method which is used to eliminate local maximums of the color components histogram is proposed. Distributing characteristic of color components histogram is analyzed quantificationally, and then the local maximums of color components histogram are confirmed and amended. The results of experiments indicate the validity of the method to eliminate local maximums of the histogram data.Using visual locating methods based on planar homography matrix estimation, the locating precision relates to the position of targets. In the application of targets locating on the movement plane of mobile robots, for each pitch degree of camera, one corresponding planar homography matrix is necessary. Aimed at these problem, a visual locating method based on planar homography estimation using RBF network is presented, this method uses the pitch degree of camera as one of the input of RBF network, makes use of the non-linear approach of RBF network fits the non-linear transformation relationship from the image plane to the movement plane, and positions the visual targets on the movement plane by planar homography estimation using RBF network. Setting different pitch degrees, the results of experiments express this method is effective in the application of target locating on the movement plane. The comparison of results of experiments shows that, under the same condition, the locating precision of this method is higher than the locating precision of the visual locating methods based on planar homography matrix estimation.On the bases of the researches on key vision technologies such as image segmentation, features extraction, and visual target locating, a way of navigation based on the guidance of path boundary is proposed. The information of the path boundary of mobile robot is attained by using visual process techniques in this method, the control parameters of navigation and obstacles avoidance are extracted from it. This navigation manner is effective for guiding mobile robot to explore the non-structure environment autonomously.更多还原