【作者】 王丽军；

【导师】 王萍；

【作者基本信息】 天津大学 ， 模式识别与智能系统， 2006， 硕士

【摘要】 台风是严重危害人类的天气之一,对台风中心准确定位具有重要意义。目前,台风中心自动定位仍处于探索阶段。台风中心定位方法之一是基于形状的定位方法。但由于台风运动过程复杂,形态多变,该方法难以适应各种情况。另一种方法是将台风等效为刚体分析其旋转性,从而确定中心。由于台风本身为非刚体,这种定位方法误差较大。无论台风在运动过程中如何变化,台风中心的运动最弱,能量最小。基于这一普遍的特点,本文提出基于小波域的嵌入式隐马尔可夫模型和交叉熵相结合的方法进行台风中心定位。该过程分以下五个步骤:1.对相邻序列图像作差分运算,提取云系的动态信息。2.对差分图像进行小波变换,获得小波系数,作为后续过程的处理对象。相对于基于像素灰度的分析,基于小波域的分析减少了观察向量的维数。3.分析单幅差分图像的空间相关性:使用采样窗口对单幅差分图像的小波域采样并将图像分割成若干个均匀分布的小块。对采样得到的小波系数进行嵌入式隐马尔可夫模型算法的训练和识别。按照空间相关性,采用嵌入式Viterbi算法、K均值算法对差分图像重新分割。4.分析差分图像序列沿时间轴的时间相关性:利用交叉熵来分析差分图像序列的时间相关性,根据图像序列的时间相关性来跟踪各差分图像序列块。较以往的基于单幅图像定位,此过程利用了多幅图像,扩大了信息利用量,而且采用空间相关性与时间相关性相结合的分析方法,能够提高定位精度。5.使用能量函数度量各序列块的运动能量,能量最小的块所对应的位置即为台风中心。与以往的方法相比,本文可提高算法的通用性。更多还原

【Abstract】 Typhoon is one of the most deadly weathers. It is significant to locate the position of typhoon center. This work is still on the study stage. One of the methods is based on shape, which is difficult to adapt to all conditions due to typhoon’s complex motion and deformable shapes. Another method is to analyze rotation feature, which deviates from the truth because typhoon is seemed as rigid body.No matter how typhoon changes in the course of moving, energy is least in the typhoon center. Based on this general feature, a new method is presented in this paper, which is by cross-entropy and embedded hidden Markov model based onwavelet-domain. It is done by five steps as follows:1. Extract the dynamic information from the differential image.2. Get wavelet coefficients of differential images. The analysis is based on wavelet coefficient, which can reduce the dimension of observation vector’s space.3. Analyze the space correlation in the single differential image: sample the wavelet coefficients with window and segment the image into the equal blocks. Train and recognize the sampled wavelet coefficients by embedded hidden Markov model. Segment the differential image into blocks over again by embedded Viterbi and K-means algorithm according to the space correlation.4. Analyze the time correlation in the differential image sequences: Use cross-entropy to track the blocks of image sequences according to the time correlation along the time axes. Compared to the old methods based on single image, the method which is based on many images can extend the used information. Moreover, it can improve the precision by analyzing the space and time correlation.5. Measure the motion energy of the blocks with the energy function and locate the position of typhoon center.This method has succeeded in extending the scope of the applied object.更多还原