实时景物理解系统研究

【作者】 刘菲；

【导师】 唐振民；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2002， 硕士

【摘要】 本课题所研究的景物理解系统作为一种智能识别系统，可以实时对周围的景物环境变化做出反应。本课题详细研究景物理解的功能，使景物理解系统能够对摄像机所拍摄的景物图像进行处理、分析、判断，从而对景物中的人类、非人类、烟雾、火等重要信息进行识别，并发出警告信息。 首先本文介绍一种实时人形识别方法，使用这个方法，可以自动监测进入摄像机镜头拍摄范围内的物体，并识别其是否是人类。其主要步骤包括，预处理其中包含了除背景之外的其它物体的图像，即，去掉背景图像，并得到该物体的形状图像，然后提取其七个不变矩，利用贝叶斯分类器做主要手段对不变矩进行分类，另外采用辅助手段修正分类结果，识别该物体的形状是否为人形，从而判断出是否有人类经过。 本课题还详细研究了对火焰、烟雾等重要的火灾信息的检测。传统的火灾报警系统一般基于红外传感器和烟雾传感器，但在大空间场合的火灾报警中失败率较高。本文利用火焰的颜色信息，火焰的图像序列的边缘不稳定和相似性等可识别特征，实现对火焰的识别。 对于烟雾的检测本系统采用边缘检测与直方图统计相结合的方式。首先对背景图像作边缘检测，并计算背景图像边缘像素占全图像素个数的百分比，然后，对序列图像的当前帧作边缘检测，并计算图像边缘像素占全图像素个数的百分比，再将两个百分比值作比，根据最后所得比值来判断是否产生大量的烟雾。 经过大量实验证明，本课题所设计的景物理解系统的检测效果很好，误判率较低，同时本系统还具有可扩展性，可以在原有的功能基础之上增加、改进各种景物理解功能，因此可以应用于很多领域。更多还原

【Abstract】 Scenery understanding system, which this paper researched as a kind of intelligent recognition system, may respond to the change of the surrounding scenery by real time. Our task is researching the function of scenery understanding in detail, and being able to process, analyze, judge the scenery’s images which are obtained by video camera, so the scenery understanding system can recognize the important information, for example, human, non human, smoke, fire, etc, and raise the alarm.First, we present a real-time human shape recognition method, which can monitor the object moving in the shooting scope of the video camera, and recognize the object being human or not. Following are the approaches used in this paper. First, we pretreat the image which include the other object besides background, that is ,getting rid of background image and obtain the shape image of the object. Second, we calculate the seven moment invariants of that image. Finally, we use Beyas assorting approach as main method assort the moment invariants and use other assistant methods to correct the result in order to recognize the human shape and judge whether human has passed or not.In this paper, we also researched the recognition of fire and smoke. The traditional fire alarming system is usually based on infrared sensor and smog sensor, but if it is in big space, the lost rate is high. This paper use effective character of the fire, for example, color information, instability and comparability of the edge of image sequence to realize the automatic alarming for fire.For the recognition of smog, the system uses the technique of integrating the edge abstracting with count histogram. First, abstract the edge of background image, and count the percentage of edge pixels in it. Second, abstract the edge of current frame image of sequence, and count the percentage of edge pixels. Finally, compare these two percentages to judge whether there is much smog or not.Many experiments proved that this scenery comprehending system is able to work well with a low lost rate. At the same time, the system is developable, that is,the system can be added or improved all sorts of scenery understanding functions based on original functions, so it can be applied in many fields.更多还原