基于视觉行为识别的煤矿安全异常监控预警基础理论及应用

Monitoring and early warning based on vision activity recognition can effectively prevent coal mine disasters. However, facing the complicated and changing underground environment, the present algorithm of activity analysis recognition cannot satisfy the requirement of mine safety monitoring with its disadvantages of  poor adaptability and strong limitation. This study, aiming at problems of coal mine safety abnormality monitoring and early warning, does research on fundamental theories based on vision activity recognition. First, it proposes a multimodal recognition algorithm of target in uncertain environmental problem domains by adopting a multicore studying method. And then, it sets up a mechanism, in which the vision activity agent is located by vision significance, on basis of the perception theory of attention mechanism. At last, it makes an analysis of the distribution of the local feature of key points of the target and explores its representing methods. This project, synthetically employing some related theories such as moving target detection, feature selection and extraction, activity recognition, etc., analyses the abnormal conditions of coal mining accident disasters by means of integral fuzzy fusion, establishes an abnormal standard feature data base for the coal mining accident disasters and constructs a multisystem coupling mechanism of underground coal mine abnormality monitoring and early warning. The results of this research have important theoretical significance and application value. For example, they enriches crucial technical theories on activity recognition, promotes further application of activity recognition technology in complex conditions and makes the management of coal mine safety monitoring and early warning more intelligent and scientific.

基于视觉行为识别的监控预警，可有效预防煤矿井下灾害的发生，但对复杂多变的井下环境，目前的行为分析识别算法，由于其适应性较差、局限性较强，无法满足矿井安全监控的要求。本项目将针对煤矿安全异常监控预警关键问题，开展基于视觉行为识别的基础理论研究。采用多核学习方法，研究不确定性环境问题领域中目标多模态识别算法；基于注意机制感知原理，建立视觉显著性定位视觉行为主体机制；分析目标关键点局部特征分布，探索关键点的局部特征表示方法。综合运动目标检测、特征选择与提取、行为识别等相关理论，通过模糊积分融合分析煤矿事故灾害的异常状态，建立煤矿事故灾害异常标准特征库，构建基于行为识别的煤矿井下异常监控预警多体系耦合模型。研究成果对丰富和完善行为识别基础理论，推动行为识别技术在复杂环境中的应用，促进煤矿安全监控预警智能化和科学化，提高煤矿安全管理水平等具有重要理论意义和应用推广价值。

本项目针对煤矿安全异常监控预警关键问题，开展了基于视觉行为识别的基础理论研究。.针对传统的马尔可夫随机场模型仅考虑像素与其邻域空间相关性的特点，提出了一种新的基于混合高斯模型的时空马尔可夫随机场的视频动态目标检测算法。通过扩展马尔可夫随机场模型，增加相邻帧像素的关联性，建立了时空马尔可夫随机场模型。进一步改进时空马尔可夫随机场的能量函数计算方法，同时考虑时间上和空间上像素相邻域势团的值，作为能量函数计算的参考项。.针对煤矿井下智能视频监控采集到的视频图像质量较差、干扰多、噪点多、目标检测不准确等问题，提出了一种基于改进粒子滤波的模糊目标检测方法。在标准粒子滤波理论框架下，以视频多图像帧差为基础构建了非线性非高斯多系统状态空间融合模型，在图像帧差得到的关键点区域范围内进行粒子抽样及概率密度的传播，用加权后验样本粒子表示多系统状态空间融合模型的后验概率密度，采用样本均值方法融合估计系统后验状态，最后对系统状态空间模型进行输出，得到目标检测结果。.研究提出了一种基于卷积神经网络的时空特征检测提取模型，该模型以多个连续视频帧为输入，从连续视频帧中提取图像特征以及时间维度上的信息，交替进行卷积和子采样操作，逐步提取出多种复杂抽象的高级特征。对多示例学习进行了扩展，提出了一种基于多示例学习的步态特征提取识别模型。.综合运动目标检测、特征选择与提取、行为识别等相关理论，分析了煤矿事故灾害的异常状态，建立了煤矿事故灾害异常标准特征库，构建了基于行为识别的煤矿井下异常监控预警多体系耦合模型。研究成果对丰富和完善行为识别基础理论，推动行为识别技术在复杂环境中的应用，促进煤矿安全监控预警智能化和科学化，提高煤矿安全管理水平等具有重要理论意义和应用推广价值。