成像目标虚拟圆特征及其应用

发布时间：2018-10-20 20:34

【摘要】：在基于成像的目标识别、定位等应用中，提取具有仿射不变性的特征是其中的关键步骤，它直接影响相关应用的最终效果。在国家自然科学基金项目“基于单目图像和方向的测距系统及关键技术研究”(608712136)和陕西省自然科学基础研究计划项目“成像目标虚拟圆特征及其应用的研究”(2011JM8002)支持下，本文对目标旋转不变特征及其应用进行深入研究，并将目标的旋转不变圆特征推广为虚拟圆特征。在综述了目前旋转不变特征提取技术及各类型不变特征量的特点基础上，依据圆形目标固有的旋转不变性质，本文提出基于多点特征的虚拟圆概念，利用图像序列中相邻图像中的3对匹配点构造了5种不同类型的虚拟圆，并进行了性能比较。与已有方法相比，这类特征易于提取，不需要对图像匹配点附加复杂的约束条件。仿真分析表明，基于正三角形扩张的虚拟圆的直径是最好的距离相关性质的线段特征，在目标距离估计中，当相邻采样时刻目标相对于相机倾斜角变化范围在[-10°,10°]之间时，应用上述线段特征进行的测距误差在±3%左右。此外，本文利用图像间4个匹配点对实现了相邻图像间目标相对姿态变化的测量，在1个像素噪声下，，测量误差不超过2.5%，具有较好的抗噪性和稳健性。
[Abstract]:The extraction of affine invariant features is a key step in imaging based target recognition and localization applications, which directly affects the final effect of related applications. Supported by the National Natural Science Foundation of China "Research on ranging system and key Technologies based on Monocular Image and Direction" (608712136) and the Research on Virtual Circle Features of Imaging Target and its Application (2011JM8002) of Shaanxi Provincial Natural Science basic Research Project. In this paper, the target rotation invariant feature and its application are studied in depth, and the rotation invariant circle feature of the target is extended to virtual circle feature. Based on the summarization of the current techniques for extracting rotating invariant features and the characteristics of the invariants of various types of features, the concept of virtual circle based on multi-point features is proposed according to the inherent rotation invariance of circular objects. Five different types of virtual circles are constructed by using 3 pairs of matching points in adjacent images in image sequences and their performance is compared. Compared with the existing methods, this kind of feature is easy to extract and there is no need to attach complex constraints to the image matching points. The simulation results show that the diameter of the virtual circle based on the expansion of the equilateral triangle is the best feature of the line segment of the range correlation property. In the estimation of the target distance, when the inclination angle of the target relative to the camera is in the range of [-10 掳, 10 掳] at the adjacent sampling time, The range error of the above line segment is about 卤3%. In addition, the relative attitude change of the target between adjacent images is measured by using four matching point pairs. The measurement error is not more than 2.5 under one pixel noise, so it has good robustness and robustness.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41;P23

【参考文献】