针对目前视频监控由于人为因素及与报警系统的被动式联动，导致误报和漏报的问题，亟需开发一种能够连续24小时实时无人监视的智能视频监控系统。该系统能够自动判别视频中的运动目标是否为人形，当某一不允许人侵入的场合或时段有人侵入时，既能实时报警又能向保卫人员准确及时发出警报。 本文针对智能视频监控系统中基于不变矩的人形识别技术进行了研究。人形识别过程包括视频帧图像提取、帧图像预处理、运动目标不变矩提取和分类识别。视频帧图像提取采用基于采样的方法，每9帧提取一帧关键帧。采得的关键帧采用背景差分去除背景后，利用大津法阈值分割得到运动目标图像，再经中值滤波去噪及Roberts算子轮廓提取得到预处理图像。然后对得到的预处理图像分别提取其Hu不变矩和Zernike不变矩。最后采用最小距离分类器和基于Parzen窗法的贝叶斯分类器进行分类，判别运动目标是否为人形。 本文建立了一个含有201个人形样本、67个动物样本的图库，并进行测试。实验证明本文基于不变矩的人形识别方法是可靠有效的。该方法对于人形的识别率可以达到85%以上。

In view of misstatements and omissions of the current video monitoring because of human factor and passive linkage for alarm system, earnestly needs to develop the intelligent video monitoring systems which can real-time unmanned monitoring for 24 consecutive hours. This system can distinguish the moving target is a humanoid or not. When someone invades a place which does not allow people to invade, the system can real-time give an alarm deterring crime, and warn to the security personnel accurately and timely. This article has researched human recognition technology based on moment invariants. The process of human recognition include video frame image extraction, image pretreatment, moving target moment invariants extraction and classification. Video frame image extraction based on sampling method, catching a frame image as a key frame from each of the 9 frames. Image pretreatment have four steps. First, remove the background of the key frames using the background difference. Second, catch moving target image using Otsu method. Third, eliminate noise by median filter. Fourth, contour extraction using Roberts operators. Then, get Hu moment invariants and Zernike moment invariants of moving target. Finally, classify the moment invariants using the minimum distance classifier and the Bayes classifier based on Parzen windows, and judge moving target whether is human. This paper established a library which contains 201 personal form samples, 67 animal samples, and texting. Proved by experiment, the human recognition method based on moment invariants is reliable and effective. The recognition rate can achieve above 85%.