【作者】 张辰；

【导师】 夏士雄；

【作者基本信息】 中国矿业大学 ， 计算机应用技术， 2014， 博士

【摘要】 运动目标检测和跟踪是运动目标行为分析的基础，是计算机感知和理解周围环境变化的重要前提，一直是计算机视觉领域中活跃的研究课题；煤炭作为我国现行使用的主要能源，煤矿安全是国家安全生产的重中之重，目前煤矿监控系统在煤矿安全生产中发挥了重要作用，而监控系统的核心是井下目标的检测和跟踪。本文针对煤矿复杂的井下环境中视频监控系统中存在的问题，在总结已有研究成果的基础上，对煤矿井下环境中运动目标的检测及跟踪中的关键问题进行研究和探讨，其主要内容如下：⒈提出了一种基于鲁棒性模糊核聚类的目标检测方法。该方法首先选取一个初始帧序列长度，利用鲁棒模糊核聚类算法将各个像素归到不同的子类中，当背景学习过程完成后，根据子类的频数选择背景保证了不同的像素对应不同的子类数，克服了混合高斯建模方法内存占用率高，计算复杂度大的缺点。⒉提出一种基于最大间隔聚类的视觉词典构造方法。该方法在利用BoW模型进行目标分类过程中，建立视觉词典库时首先用K‐均值得到初始聚类值，然后使用最大间隔聚类进行优化，有效提高了视觉词典的精确度，分类效果得到了明显的改善。⒊提出一种基于极速学习机和多特征融合的粒子滤波跟踪方法。该方法采用极速学习机建立粒子与粒子在后验分布中概率大小的权值两者之间的函数关系，根据这个函数重新调整粒子的权值，从而有效避免粒子的贫化。同时，考虑到粒子滤波框架下单一视觉特征对目标外观变化表示不充分的缺点，利用全局颜色特征和局部DAISY特征融合构建目标表示模型，提高了目标跟踪算法的鲁棒性。⒋提出一种结合协同训练分类器的粒子滤波跟踪方法。该方法为了有效抑制分类器学习中出现的误差积累导致的漂移现象，引入协同训练更新分类器，提高检测精度；检测器与粒子滤波跟踪器同时运行，然后将跟踪结果和检测结果进行融合得到最终跟踪结果。利用无重叠分类器网格解决跟踪失败后粒子滤波重启跟踪问题。更多还原

【Abstract】 Detecting and tracking moving object is the basis of behavioral analysis ofmoving targets, and has always been an active research topic in the field of computervision. The video surveillance system currently plays an important role in the safetyproduction of coal mine. Against the problems in the video surveillance system underthe complex environment of coal mine, this dissertation studied and discussed the keyproblems in the detecting and tracking of the moving object in the coal mine bysummarizing the existing research results with its main contents being as follows:1. Against the characteristic that target and background change dynamically, thisdissertation proposed a moving object detection method based on robustness FuzzyKernel-clustering. Then we could overcome the disadvantages of high computingcomplexity and high memory usage rate of the MOG method.2. Against the disadvantages of sensitive starting point and unstable clusteringresults caused by K-mean algorithm based on the BoW model, this dissertationproposed the EMMC method. With the EMMC method the initial clustering obtainedthrough the K-mean in the BoW model was optimized to effectively improve theaccuracy of the visual dictionary.3. For the particle degradation in the particle filter algorithm in the process oftracking the moving object, this dissertation proposes the ELM method to improve theparticle filter tracking. Meanwhile, we used the global color feature and local DAISYfeature fusion to build the target representation model and improve the robustness ofthe target tracking algorithm.4. Against the problem that particle filter cannot be applied to the long-termtracking occasions, this dissertation proposed a long-term particle filter trackingmethod that combines the co-training classifiers. The non-overlapping classifier gridwas applied to solve the problem of restarting the particle filter tracking after trackingfailure.更多还原