| 题名 | 智能视频监控系统中若干关键技术研究 |
| 作者 | 毕国玲 |
| 学位类别 | 博士 |
| 答辩日期 | 2015-05 |
| 授予单位 | 中国科学院大学 |
| 导师 | 续志军 |
| 关键词 | 智能视频监控 图像增强 目标检测 特征匹配 目标跟踪 |
| 其他题名 | Research of Several Key Techniques in Intelligent Video Surveillance System |
| 学位专业 | 机械电子工程 |
| 中文摘要 | 近年来，公共安全形势变得愈加复杂和严峻，使得视频监控规模日益扩大，其监控方式也从传统的视频监控方式向智能视频监控方向发展。然而，智能视频监控技术是计算机视觉、模式识别、人工智能、数据挖掘等多学科的交叉和结合，且面临的问题和实际应用场景复杂，至今仍处于探索发展阶段。 在实际的智能视频监控系统中，由于恶劣天气（雾霾、雨、雪等）、夜晚低照度、光照不均匀等环境造成的图像质量下降，给后面的视频分析带来了先天的困难；动态背景、光照变化、相机抖动等复杂的背景导致背景模型建立及前景目标分割困难；面对海量的视频图像数据，需要准确地检索或查询相关目标信息，使得对鲁棒的目标特征提取及匹配技术的需求也愈加的急迫；对目标尺度变化、发生遮挡、目标颜色与背景相似等复杂环境中运动目标的长期稳定跟踪也是一个亟待解决的问题。对于上述问题，目前的智能视频监控相关技术仍不成熟，易导致虚报和漏报现象的发生，从而制约了智能视频监控系统实战性能的提升。本文针对智能视频监控系统中图像增强、复杂背景建模及前景目标检测、目标特征提取及匹配、目标跟踪等关键技术进行了深入地研究。主要完成的工作总结如下： 1. 结合人类的视觉特性和广义对数比模型，提出了一种多尺度图像增强算法。根据视觉系统的全局自适应调节特性，对图像作全局亮度的类对数变换；利用主观亮度感觉与实际光强对数的局部线性关系以及视觉敏感特性，结合四方向Sobel梯度图像，调节图像的局部对比度；采用自适应的不同尺度引导滤波结合广义对数比模型将不同尺度图像的有效信息进行融合，得到最后的多尺度增强图像。试验证明本算法实现了图像对比度提高和动态范围的有效压缩，增强和保留了图像细节，具有较强的抗噪能力，有效解决了视频监控系统中可见光低照度图像、红外图像等降质图像的增强问题。 2. 根据Retinex理论中的照射_反射模型提出了一种图像增强算法。算法采用具有边缘保持特性的不同尺度引导滤波核函数作为环绕函数，来估计反映图像整体结构的低频照射分量；所有运算采用广义对数比模型的有界运算代替传统的线性运算，去除照射分量，将不同尺度的反射分量从原始图像中分割出来；通过对不同尺度反射分量的有效融合，最终得到反映物体本质特性的多尺度反射分量的增强图像。试验结果表明算法避免了传统多尺度Retinex算法中的光晕伪影现象以及传统运算中的越界问题。此算法尤其对夜间图像、雾霾图像的细节增强和对比度及清晰度的提高效果明显。 3. 背景建模及前景目标检测方面：1）为了对运动状态发生变化的目标进行检测，提出了一种基于双背景模型的目标检测方法，实现了视场内遗留物和搬移物的高效检测；2）在分析和总结ViBe算法特性及不足的基础上，提出了一种基于随机聚类的复杂背景建模及前景目标检测算法。算法充分考虑同一位置的像素在时间上的关联性和与其相邻像素的空间关联性，有效地提高了背景模型对动态背景、光照变化及相机抖动等复杂背景的适应性和鲁棒性，实现了对前景目标的准确检测。 4. 提出了一种基于角点检测和灰度差分不变量（GDI）的局部特征描述子的快速匹配算法。算法通过建立较少图层的阶梯金字塔，利用Shi-Tomasi算法提取图层中数量适中的强角点，采用具有几何意义的低阶GDI建立其局部特征区域统计直方图的描述子，最终实现角点的稳定匹配。试验证明，算法具有旋转、尺度缩放、光照及小视角变化、模糊等性能不变性，其匹配实时性和匹配精度均较高。 5. 提出了一种基于颜色特征和纹理特征相融合的Mean Shift目标跟踪算法。改进Mean Shift算法中的Epanechnikov核函数模型，降低了计算量并突显跟踪窗口内目标灰度信息；考虑了目标跟踪的尺度的自适应性。从而使得目标尺度发生变化、目标和背景相似的情况下，也能很好地实时跟踪运动目标。 |
| 英文摘要 | In recent years, public security situation has become increasingly complex and serious, making the scale of video monitoring increased, and the monitoring mode develop from the traditional video monitoring to intelligent video surveillance. However, intelligent video surveillance technology is a multi-disciplinary crossover and combination involved with computer vision, pattern recognition, artificial intelligence, data mining, and so on, the facing problems and practical scene are complex, and it is still at the stage of exploration and development. In actual intelligent video surveillance system, complex environment such as bad weather (haze,rain,snow), lack light of dark night, uneven illumination, and so on, causing the loss of image quality, which bring the innate difficulties to the subsequent video analysis; complex background such as dynamic background changes, illumination changes, camera shakes, making background modeling and foreground detection difficult; in the face of huge amounts of video image data, it is necessary to retrieve or query related target information accurately, making the demand of robust target feature extraction and matching technology become increasingly urgent; long-term stably tracking of video moving target in the circumstances of scaling, occlusion, object resembles the background, is also an urgent issue to tackle. For the above mentioned problems, the current intelligent video surveillance and related technology is still immature, easily lead to adverse consequences of misrepresentation and omission, which restricted the performance of the intelligent video surveillance system in practical applications. This article focuses research on several key techniques in the intelligent video surveillance system, such as image enhancement, complex background modeling and foreground object detection, feature extraction and matching of target, target tracking. Main work is summarized as follows: 1. A multi-scale image enhancement algorithm is proposed which combining visual properties and general log-radio model. The algorithm uses global adaptive adjustment of the human visual properties and takes a similar logarithmic transformation to the global image brightness; using local linear relationship between the human subjective feelings and the actual light intensity and the sensitive characteristics of human visual, combines with four direction Sobel gradient image, to adjust local contrast of the image; using the adaptive different scales of guide filter function and the generalized log-ratio model, integrating effective information of different scales images to get the final multi-scale enhancement image. The experimental results show that the proposed algorithm has realized the image contrast enhanced and the effective dynamic range compressed, strengthened and kept the details of the image texture and edge, with a stronger anti-noise ability, effectively solved the problem of enhancement with low illumination image and infrared image in video monitoring system. 2. According to Irradiation_Reflection model of Retinex an image enhancement algorithm is proposed. Using the edge-preserving and adaptive guide filter function as the surround function to estimate the different scales irradiation images which react the whole structure of image; using the bounded generalized log-ratio model instead of traditional operation, removing illuminate components from the original image to segment the different scales of the reflection image; fusing the effective information of the different scales reflection images and getting the final multi-scale reflection enhanced image which react the nature of the object. The experimental results show that the proposed algorithm overcomes the emergence of halo effect of the multisacle Retinex and computing overflow effectively. The effect of the algorithm is particularly obvious for night images and haze images. 3. In the respect of complex background modeling and foreground object detection: 1) in order to detect the change of movement state of target, propose an efficient method for foreground objects detection based on dual background models, realizing the detection of abandoned and moved objects in the field of view; 2) on the basis of the analysis and summary of the features and disadvantages of ViBe algorithm, a kind of complex background model and foreground detection method is proposed. The algorithm gives comprehensive consideration to the same location pixels of the relevance of time and the correlation of space with its adjacent pixels, which can significantly improve the adaptability and robustness of the background model such as dynamic backgrounds, illumination changes and camera shakes, achieving the goal of accurate detection of foreground. 4. A fast matching algorithm is proposed based on corner detection and Gray-value Differential Invariants (GDI) local feature descriptor. Establish less layer pyramid and use Shi-Tomasi algorithm to extract a reasonable number of strong corners for the layers. Using low-level gray scale differential invariants with geometric meaning to establish a local feature descriptor based on region histogram, realizing the stability of the corner matching. Experimental results demonstrate that the proposed algorithm has the invariance for rotation, scaling, illumination changes, smaller viewpoint changes, blur and so on, the algorithm has better matching prec- ision and real time. 5. A Mean Shift target tracking algorithm is proposed based on the fusion of color features and texture features. Improves the Epanechnikov kernel function model of the Mean Shift algorithm, reduces the amount of calculation and highlight the target gray level information inside the tracking window; considering the adaptability of the scale with target tracking. When the scale of target changes, the target and background are similar, can also be a good real-time tracking for moving targets. |
| 公开日期 | 2015-12-24 |
| 内容类型 | 学位论文 |
| 源URL | [http://ir.ciomp.ac.cn/handle/181722/48824]  |
| 专题 | 长春光学精密机械与物理研究所_中科院长春光机所知识产出 |
| 推荐引用方式 GB/T 7714 | 毕国玲. 智能视频监控系统中若干关键技术研究[D]. 中国科学院大学. 2015. |
| 个性服务 |
| 查看访问统计 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
修改评论