【作者】 张谢华；

【导师】 张申；

【作者基本信息】 中国矿业大学 ， 信息与通信工程， 2013， 博士

【摘要】 目前我国大多数的煤矿视频监控系统还主要停留在人工监控阶段，智能化煤矿视频监控系统是发展的必然趋势。它可以自动采集获得视频监控图像序列，进行实时运动目标检测、识别和跟踪，通过理解分析图像画面主动发现违规行为、可疑目标和潜在危险，以快速合理的方式发出警报，指导启动相应的联动控制措施。煤矿智能视频监控系统的实现，需要综合运用图像处理、机器学习和计算机视觉等领域中的多项技术，本文对其中的四类关键技术进行研究，具体工作包括：为了对伴有随机噪声的煤矿雾尘图像进行清晰化处理，提出一种基于DCPBF的去雾除尘和同步去噪算法。推导建立煤矿雾尘降质图像退化模型；设计基于暗原色先验知识的环境光、粗略透射率估计方法与步骤；采用联合双边滤波快速获得精细透射率图；依据图像退化模型构建正则化目标函数，求取转换图像并进行高斯双边滤波，获得去雾除尘图像且同步实现噪声的有效去除。针对相对静止的煤矿视频监控环境背景，采用背景减除法进行运动目标检测。提出基于聚类技术的自适应背景建模与更新方法，利用改进的FCM算法对像素灰度取值进行聚类，自适应选取不同个数的聚类构建各像素背景模型，随场景变化进行聚类修改、添加和删除完成背景更新。联合背景差分信息、三帧差分信息和空间邻域信息进行前景检测，通过改进的OTSU方法自动设置差分阈值。提出结合像素亮度和纹理特征的运动阴影检测方法，依据在阴影覆盖前后的灰度图像中，像素具有亮度值相关性和纹理特征值不变性，实现运动阴影的检测与去除。将单目标跟踪看作为目标和背景的在线分类问题，选用线性SVM作为分类工具，提出一种添加样本约简机制的FLSVMIL方法实现分类器在线更新，并提出基于FLSVMIL的单目标跟踪算法。由于可能受到无效历史信息的干扰，并且难以处理样本集非线性可分的问题，提出基于LSVMSE的单目标跟踪算法，采用集成分类器进行运动目标跟踪。根据煤矿智能视频监控系统中多目标跟踪的任务需求，提出基于UKF-MHT的多目标跟踪算法。设计算法的基本框架，确定关键步骤的处理方法，其中包括跟踪门设置、目标预测值与观测值的数据匹配、航迹评价与删除、航迹聚类和m-best假设的产生以及目标状态的预测更新。在自适应跟踪修正阶段，针对由目标短暂丢失、粘连和分裂可能引起的三类跟踪错误，设计具体的判别策略和修正方法。更多还原

【Abstract】 At present most of our coalmine video surveillance system is still at the stage ofmanual monitoring, intelligent system is an inevitable trend of development. It canautomatically capture surveillance video image sequence, detect, identificate andtrack moving targets in real time. By analyzing image screen, the intelligent systemcan proactively uncover violations, suspicious targets and potential hazards.And thenquickly alerting by a reasonable way, it can guide the activation of correspondinglinkage controls. The implementation of coalmine intelligent video surveillancesystem requires using a number of technologies such as image processing, machinelearning and computer vision and so on. This paper studies four key technologies; itsspecific tasks are as follows:In order to clear coal fog dust images with random noise, an algorithm of fogdust removal and simultaneously denoising based on DCPBF is proposed. This paperestablishes a coalmine fog dust image degradation model, designs methods andprocedures for estimating ambient light and rough transmittance based on darkchannel prior. The fine transmittance diagram is quickly obtained by joint bilateralfiltering. A regularization objective function is constructed based on the imagedegradation model. By solving a converted image and Gaussian bilateral filtering theimage, fog dust removal and simultaneously denoising are realized.Aiming at the relatively static background of coalmine video surveillanceenvironment, this paper uses background subtraction method for moving targetdetection. An adaptive background modeling and updating method based onclustering technology is proposed. Clustering pixel gray values by improved FCMalgorithm, a different number of classifications are adaptively selected to build thebackground model of each pixel. With scene changes classifications are updated,added and deleted thus completing background update. Image foreground is detectedby jointing background differential information, three differential information andspatial information. Differential threshold is automatically setted by improved OTSUmethod. A moving shadow detection method using pixel luminance and texture isproposed. Because the pixel brightness and texture value are invariant in gray imagesbefore and after shadow covering, moving shadow detection and removal can berealized.Considerring single-target tracking as an online classification problem between target an background, a linear SVM is used as a classification tool. An FLSVMILmethod with sample reduction mechanism is proposed to realize online updatingclassifiers. And so an single-target tracking algorithm based on FLSVMILis proposed.Due to the interference of invalid history information and the nonlinear separability ofsample set, an single-target tracking algorithm based on LSVMSE is proposed.Moving target is tracking by an ensemble classifier.According to the requirement of multi-target tracking mission in the coalmineintelligent video surveillance system, an multi-target tracking algorithm based onUKF-MHT is proposed. This paper designs the basic algorithm framework, anddetermines the treatment of critical steps which include setting tracking gate,matching target predicted value and observed value, evaluating and removing track,clustering track and generating m-best hypotheses, predicting and updating targetstates. In the process of adaptive tracking correction, specific discriminant strategiesand correction method are designed for three types of tracking error caused by targettemporary loss, target adhesions and split.In this paper, there are fifty-seven figures, twenty-one tables, and one hundredand fifty-eight reference documents.更多还原