【作者】 卓志敏；

【导师】 杨莘元；

【作者基本信息】 哈尔滨工程大学 ， 通信与信息系统， 2007， 博士

【摘要】 在高新技术兵器迅速发展的当今世界，基于红外探测的各类武器系统已成为世界各国重点研究和发展的项目之一。红外成像目标检测与跟踪技术一直是困扰和制约红外成像系统实用性能的瓶颈问题和技术难点而亟待解决。本文对红外成像目标检测与跟踪问题进行了深入研究，研究内容主要包括红外弱小目标检测、红外成像运动目标检测、红外成像全局运动估计以及红外成像目标跟踪。根据红外目标与背景边缘对不同结构元素感知能力的不同，提出了一种基于多结构元素灰度形态学背景估计的红外弱小目标融合检测算法，该算法首先采用四种不同方向的线性结构元素分别对图像进行Top-Hat变换，然后对变换结果采用“乘”融合处理来突出目标抑制背景，最后对融合结果进行自适应阈值分割。它克服了传统基于单结构元素的处理算法对图像适应性不强的弱点，更有利于增强目标与抑制背景。根据红外成像运动目标的检测要求，提出了一种结合帧差分法和背景差分法的运动目标检测算法。该方法首先利用图像匹配算法对背景图像进行匹配，再利用得到的匹配结果进行全局运动估计和补偿，之后基于帧差分方法获得目标运动区域，最后对多帧图像得到的运动区域进行比较，并在运动区域无重叠的时候利用多帧图像信息对当前帧图像进行背景重建，并对重建后的背景运用背景差分法进行运动目标检测。该算法有效地解决了帧差分法对运动目标检测时容易出现“空洞”的问题，其最大优点是能够检测出完整的运动目标。针对RANSAC算法仅仅利用部分内点信息的不足，提出了一种基于RANSAC+LS算法的红外成像全局运动估计方法。该算法首先对待匹配块进行预检验，然后利用RANSAC算法得到尽可能多的内点，最后利用LS算法对这些内点进行运动参数估计。该算法计算速度快，比传统的RANSAC算法利用了尽可能多的内点，并能克服传统LS算法易受错误匹配结果影响的缺点。本文将提出的算法和LS估计、M估计、LMedS算法及RANSAC算法进行了性能比较，比较结果表明本文算法估计精度高于以上几种算法。针对非线性非高斯情况下的目标跟踪问题，提出了一种基于UPF算法的红外成像目标跟踪算法，该算法利用UKF滤波得到一个参考分布，然后利用这个参考分布来代替经典PF算法中的参考分布p(x_k|x_k-1)作为重要性密度函数，该算法克服了粒子滤波应用过程中经常可能出现的退化现象以及选用p(x_k|x_k-1)作为重要性密度函数时没有利用最新观测数据的缺点。本文将UPF算法与EKF算法、UKF算法及PF算法进行了性能比较，仿真结果表明UPF算法解决非线性滤波问题的能力明显优于以上几种算法。本文还将UPF算法应用到了红外成像目标跟踪问题中。综上所述，本论文对红外成像目标检测与跟踪相关技术进行了深入的研究，对提出的几种算法均利用实拍的红外图像进行了试验验证，试验结果表明本文提出的算法获得了很好的检测与跟踪效果。更多还原

【Abstract】 Nowadays, with the development of high-tech weapon, a kind of weapon system based on infrared detection becomes one of the key studying and developing projects in various countries. Infrared target detection and tracking technology is always the bottleneck problem that affects the performance of infrared equipment, and the problem must be solved urgently. The research and analysis on this problem is proposed deeply and extensively in this paper. The research content mainly includes the infrared dim target detection, infrared image moving target detection, infrared global motion estimation and infrared target tracking technology.According to the different awareness between infrared target and background edge with different structuring elements, an infrared dim target fusion detection method based on multiple structuring elements background estimation is proposed in this paper. Firstly, this method processes the infrared image using Top-Hat with four linear structuring elements. Secondly, this method fuses the first step result using multiplication to stand the object out of the background. Finally, this method detects the target using adaptive threshold segmentation. This method overcomes the bad picture compatibility of the traditional method based on single structuring element, and this method has good performance in goat enhancement and background suppression.A method of infrared image moving target detection in a complex environment is proposed in this paper according to the request of infrared system. This method fuses the frame subtraction method and the background subtraction method. Firstly, this method matches background and makes the motion compensation, and then the motion region based on frame difference can be obtained. Secondly, by comparing with the multi-frames motion region, this method reconstructs the background when the motion region has no overlaps. Finally, the target can be detected by background subtraction method. Thecomplete mdtion target can be detected with this method.A method of global motion estimation based on RANSAC+LS algorithm is proposed in this paper. Firstly, this method previews the matching block and then acquires enough inliers using RANSAC algorithm, secondly, it estimates the global motion estimation parameter using LS algorithm. This method has high efficiency and this method uses inliers as many as possible compared with traditional RANSAC algorithm, and it can overcome traditional LS algorithm’s disadvantage. This paper compares RANSAC+LS algorithm with LS estimation method, M estimation method, LMedS algorithm and RANSAC algorithm. The simulation result shows that the RANSAC+LS algorithm has better performance than traditional algorithms.To solve the target tracking problem in non-linear and non-gaussion noise, an infrared target tracking method based on UPF algorithm is proposed in this paper. This method generates proposal distributions using UKF instead of the transition prior p(x_k|x_k-1) as the proposal distribution. Comparing with traditional PF algorithm, this method solves the degeneracy and overcomes the disadvantage when choose the p(x_k|x_k-1) as the proposal distribution. Experiments are done to compare the performance of the four algorithms: UPF, EKF, UKF, and PF, and the results shows that the new particle filter outperforms obviously the standard particle filter and other filters such as EKF and UKF. This method is used in infrared image tracking.In summary, the infrared target detection and tracking problems are researched in this paper, and new algorithms have been proposed. The experimental results indicate that the algorithms can get better results in the target detection and tracking field.更多还原