云层背景下目标多特征信息融合及跟踪策略研究

	云层背景下目标多特征信息融合及跟踪策略研究
	郭同健
学位类型	博士
导师	高慧斌
	2014-07
学位授予单位	中国科学院大学
学位专业	机械电子工程
摘要	靶场光电跟踪测量系统是利用光学成像技术跟踪测量空间的飞行目标，以获取目标的高精度测量数据，为武器、航天试验鉴定及故障分析提供重要依据。近年来，新的试验任务需求和靶场建设发展需要对靶场光电跟踪测量系统连续稳定跟踪目标的能力提出更高要求。然而，大量外场试验结果表明当前靶场光电跟踪测量系统的目标跟踪技术存在着对天气条件过于依赖的缺陷。尤其是在晴空中夹杂许多云层的情况下，一方面极易将云层错误识别成目标导致跟踪中断；另一方面在目标被遮挡后缺乏高效的搜索再跟踪能力，大多数需要花费较长时间并且依赖外界引导数据及人员的判决操作进行目标搜索再跟踪。由于晴空夹杂云层是靶场试验任务中常见的气象条件，为了避免局部云层的影响而错失后续的测量数据甚至是关键段的珍贵数据，极有必要研究并解决云层背景下目标跟踪易中断的问题。本文以靶场光电跟踪测量系统为研究平台，以提高云层背景下目标跟踪准确性和平稳性为目的，围绕当前云层背景下目标跟踪方法存在的不足之处，开展目标识别、目标跟踪及目标搜索相关关键技术的研究。在方法具有实用价值的指导思想下，提出靶场光电跟踪测量系统在云层背景下稳定跟踪的实现方案。全文完成的主要工作归纳如下：研究了云层背景下疑似目标区域分割技术。图像分割可以有效降低后续目标识别的运算开销，针对现有的单一图像分割技术难以实现云层背景下疑似目标区域的有效分割，确立依据跟踪模式的特点采用相适应的图像分割技术思路。本文将靶场光电跟踪测量系统的目标跟踪过程划分成稳定跟踪模式和主动搜索模式。在稳定跟踪模式下，依据目标大致位置区域已知的特点，采用改进后的最大类间方差阈值分割方法进行快速图像分割；在主动搜索模式下，依据全视场分割的要求，提出基于视觉注意机制的图像分割方法实现精准的疑似目标区域分割，通过算例验证了方法的可行性。研究了目标多特征信息融合识别技术。针对目标周边存在云层或被云层遮挡时，采用单一的灰度特征易将云层错误识别成目标的问题，本文提出了基于神经网络和证据理论两级联合的多特征信息融合目标识别方法。首先提取分割后疑似目标区域的多种特征信息，对每种特征信息通过相应神经网络进行特征识别，再利用证据理论对多个神经网络的识别结果融合判决。该方法一方面利用了目标多特征信息降低目标误识别的概率，另一方面两级联合的信息融合模型体现出良好的可扩展性，能方便地根据不同情形、不同对象增减改变特征融合信息源，具有很强实用价值。实验结果证明该方法能准确辨别出云层和目标。研究了目标运动状态估计反馈跟踪技术。针对传统的目标脱靶量直接反馈跟踪方式难以适应云层背景下目标跟踪需求的问题，本文深入研究了在靶场光电跟踪测量系统中采用目标运动状态估计反馈的跟踪方法及实现细节。分析了不同应用情形下目标运动状态估计的可观测性问题，探讨卡尔曼滤波、不敏卡尔曼滤波及交互多模型滤波三种滤波估计方法的应用改进，给出实际应用时目标运动模型建立、数据时空对齐、数据关联及状态估计等实现过程。实验结果表明该方法弥补了直接反馈跟踪方式的跟踪性能严重依赖于测量数据质量的缺陷，提高了云层背景下目标跟踪稳定性。研究了运动目标主动搜索策略。针对目标跟踪过程中受到云层遮挡而短暂丢失目标的情形，提出了基于最大概率区间的目标主动搜索方法；针对长时间丢失目标且缺乏外界引导数据的情形，提出了在拦截点处采用警戒线目标搜索方法来提高搜索效果，同时针对警戒线搜索过程中目标成像会出现运动模糊的情况，给出了运动模糊恢复方法。实验结果表明主动搜索增大了目标有效搜索区间，可以大幅提高目标搜索成功概率。综上，本课题研究创新之处在于：通过对靶场光电跟踪测量系统现有的目标跟踪技术研究和改进，提出并验证了在云层背景下工作时目标多特征信息融合及稳定跟踪的相关关键技术，该技术的意义在于解决了云层背景下目标跟踪易中断的问题，弥补了当前靶场光电跟踪测量系统的目标跟踪技术对天气条件过于依赖的不足，在目标跟踪测量领域具有一定的实用价值。
其他摘要	A range photoelectronic tracking measurement system is the use of optical imaging techniques tracking measurement space flight target, in order to get high precision measurement data of the target, and provide important basis for test evaluation and fault analysis of weapons and spaceflight. In recent years, the needs of new test mission and range construction development put forward higher requirements for range photoelectric tracking measurement system to continuously and stably track target. However, large numbers of field test results show that the current target tracking technology of electro-optical tracking measurement system exists defects that it relied too much on the weather conditions, especially the inclusion of many clouds in the clear sky. On the one hand, it is easy to misidentify clouds into the target which will leads to tracking interruption; on the other hand, it lacks the ability of effective searching and tracking again when the target get covered, most of the time it takes a long time and rely on external guide data and personal judgment operation to search and track target again. The clear sky with clouds is a common meteorological condition in shooting range test task. To avoid missing out on the follow-up measurement data and even the key period of valuable data for the influence of partial clouds, it is necessary to study and solve the problem of the easily interruption with tracking under the background of clouds. The paper is based on the range photoelectric tracking measurement system as the research platform, on the purpose of improving the target tracking accuracy and stability as well as around the deficiency of current target tracking method under the background of the cloud, to carry out the relevant key technologies research of target identification, target tracking and target search. With the ideology of practicality, the stable tracking scheme is proposed under the background of clouds.The main works accomplished in the paper summarized as follows: The suspected target region segmentation technology is researched under the background of the clouds. Image segmentation can effectively reduce the operation costs of subsequent target identification. In view of the existing single image segmentation technique is difficult to realize effective segmentation suspected target area under the background of clouds, an idea is established that image segmentation technology is adopted flexibly according to the characteristics of the tracking mode. The target tracking of range photoelectric tracking measurement system is composed with stable tracking mode and active searching mode. In stable tracking mode, beacause the rough target location is known, the improved maxinum between-cluster variance threshold segmentation is adopted for fast image segmentation. In active searching mode, due to the requirements of the full view-field segmentation, the image segmentation method based on visual attention mechanism is proposed to realize precise suspected target area segmentation. The experimental results show that the method is feasible. The target multi-feature information fusion recognition technology is researched. If the target is surrounded or covered by clouds, the cloud is easily identified into target through the single gray characteristics. Thus, the multi-feature information fusion recognition method is proposed basing on two-level jointing neutral networks and evidence theory. Various feature information of the segmented suspecting target area is extracted firstly, and then the corresponding neutral network is used to feature recognize. Afterwards, evidence theory is utilized to incorporate multiple neural network outputs and make the synthetical decisions. On the one hand, the target multi-feature information is used to improve recognition accuracy. On the other hand, the two-level joint information fusion model shows well extensible. It has strongly practical value that it can easily increase or decrease feature fusion information source according to different situations or different object. The experimental results show the clouds and the target can be accurately identified through the method. The target motion state estimation feedback tracking technology is researched. Because the traditional way of direct feedback tracking target is difficult to meet the requirement of target tracking under clouds background, the target motion state estimation feedback tracking technology is studyed deeply in the range photoelectric tracking measurement system. The observability problem is analyzed in different application situations. Kalman filter, unscented kalman filter and interactive multiple model filter are discussed and improved for the application. The target motion model, data space and time alignment, data association and state estimation are achieved. The experimental results show that this method makes up for defects that the target tracking performance of direct feedback seriously dependent on the measuring data quality, and improves the stability of target tracking in the cloud backgrounds. The active searching strategy of moving target is studied. It puts forward the target active searching method based on the maximum probability areas in the situation of cloud-covering and temporary loss of target. In order to speed up the searching results for long losting goals and lack of external guide data, it puts forward the warning line target searching method at the intercepting point. Because the image motion blur appears in the warning line target searching method, the blur motion recovery method is also proposed. The experimental results show that the active searching increases target searching range, and effectively improve the probability of target successful acquisition. To sum up, the innovations of the paper lie in: through the researching and improving of existing techniques in the range of photoelectric tracking measurement system, the related key technologies of multi-feature information fusion and stable tracking is put forward when working at the background of clouds. The significance of the technology is to solve the problems of easily tracking interruption under the background of the clouds, to make up for the defect of current range target tracking technology which relyes too much on the weather. The method has certain practical value in the field of target tracking measurement.
语种	中文
文献类型	学位论文
条目标识符	http://ir.ciomp.ac.cn/handle/181722/41411
专题	中科院长春光机所知识产出
推荐引用方式 GB/T 7714	郭同健. 云层背景下目标多特征信息融合及跟踪策略研究[D]. 中国科学院大学,2014.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
郭同健.pdf（2114KB）			开放获取	CC BY-NC-SA	请求全文