【作者】 蒋亮； 张翔； 陶霖密； 徐光；

【Author】 JIANG Liang1+, ZHANG Xiang2, TAO Lin-mi2, XU Guang-you2 1(School of Software, Tsinghua University, Beijing, 100084, China) 2(Department of Computer Science and Technology, Tsinghua University, Beijing, 100084, China)

【机构】 清华大学软件学院； 清华大学计算机科学与技术系；

【摘要】 普适计算大环境下的计算机视觉处理包含运动检测和前景提取,目标跟踪,行为(事件)分析,现场存档和及时报警等多个相辅相成的步骤。前景目标的检测和提取是其中的第一个步骤。目前的运动检测和前景提取方法主要有时域模板法,高斯混合模型法,非参数核密度估计方法,光流法,Wallflower,贝叶斯分类法等等,但是它们都假设出现频率最高的像素亮度值为背景亮度值,从而无法对实际前景目标进行持续地检测。实验证明,这些方法不适合室内环境下大前景目标各部分运动状态不一致情况下前景的完整持续检测,也不能够对室外监控环境下较小前景目标的运动状态变化较大时对其进行持续地检测,从而给后续目标分类和跟踪,事件检测等处理步骤造成很多限制和困难。本文基于背景减法和帧间差法相融合的思路,提出了双背景模型及其更新机制和前景的四值表示法。实验结果表明本文双背景模型及其更新机制可以适应光线的变化并去除阴影的影响,较好地避免噪声,并在很短时间内自动去除很少情况下较小区域的误检和漏检,得到了干净的背景图,从而对不同场景下的运动目标进行了持续地检测;前景的四值表示法可以将前景目标的运动状态比如运动前景部分,静止前景部分,稳定前景部分和背景建模至前景图中,以便后续进一步跟踪和行为分析的处理。更多还原

【Abstract】 Pervasive computing has become an important research area and computer vision as a front end has more and more applications. Motion detection or foreground extraction is the first step of a typical vision system and has been researched extensively. Gaussian mixture model (GMM), Non-parametric kernel density estimation (NPKDE), optical flow, temporal templates, Bayesian classification and Wallflower are traditional motion detection techniques that are designed to adapt to different situations. All these techniques are based on the assumption that the most frequently observed pixel values are background values. However, this assumption is not valid in many scenarios and will put constraints on latter processing such as tracking or event detection. For instance, in the cases of parking lots and meeting rooms, objects change their kinetic states time to time. All the above motion detection techniques will update the slow motion objects, which is the true foreground, into the background, which results in partial or even no detection of the foreground objects. In this paper, an algorithm is proposed based on background subtraction and inter-frame differencing. Two background models, the original loaded background and the runtime background, are created and dynamically updated in whole detection process. Relatively a novel four-value representation is used to describe the detection results: moving foreground, static foreground, stable foreground and background, based on the two models. Motion detection experiments in meeting rooms and building halls demonstrated the algorithm can consistently detect the foreground throughout the video sequences without detecting the illumination change or shadows. The experimental results also show that the algorithm can get the clean background and integral foreground based on the updating mechanism.更多还原