热红外视频的异常行为检测

发布时间：2018-09-08 20:27

【摘要】：在经典的计算机视觉领域,由于传统的色彩视频包含信息复杂,再加上由于光照变化、烟雾等不利因素的影响,使得基于传统彩色视频的异常行为检测的效果不好,帧率很低。而红外热像仪热红外视频由于包含信息单一,并且在夜视、烟雾和光照剧烈变化的复杂环境下仍能有效工作。因此,在本文,我们使用FLIR FC-S系列红外热像仪采集热红外视频,并基于采集的热红外视频来进行异常行为检测。本文的主要研究内容为:1、本文利用基于轮廓的背景减除算法,提取热红外视频中的前景目标。该方法的主要步骤为:使用基于统计的背景减除方法提取出感兴趣的区域,然后建立轮廓显著图并进行细化;使用分水岭变换方法补全轮廓线;最后使用FloodFill方法填充轮廓并得到精确的前景目标。2、本文提出了一个基于单元格的异常行为检测算法。由于异常行为发生概率低、持续时间短而且无周期性,因此,很难对异常行为进行建模。而本文采取的策略为:先对训练集和测试集的视频提取前景,然后将这些前景划分为网格,并对每个单元格提取局部特征,最后基于训练集的特征构建出正常行为模型。然后,将测试集上每个单元格的特征与正常行为模型进行对比,如果相近,就判定为正常行为,否则为异常行为。最后,本文建立了十四段包含不同异常的视频集,并在其上进行测试。实验结果表明:本文提出的方法有效地对发生的异常行为进行了识别和跟踪,而且该算法在检测异常时,能将帧率保持在11fps左右,基本可以满足日常的监测工作。
[Abstract]:In the classical computer vision field, because the traditional color video contains the information complexity, plus because of the illumination change, the smoke and other adverse factors influence, makes the traditional color video abnormal behavior detection effect is not good, the frame rate is very low. Thermal infrared video of infrared thermal imager can work effectively in complex environment of night vision, smoke and illumination because of its single information. Therefore, in this paper, we use FLIR FC-S series infrared thermal imager to collect thermal infrared video, and based on the collected thermal infrared video to detect abnormal behavior. The main content of this paper is 1: 1. In this paper, the foreground target of thermal infrared video is extracted by using contour-based background subtraction algorithm. The main steps of the method are as follows: the region of interest is extracted from the background subtraction method based on statistics, and the salient contour map is built and refined; Finally, we use the FloodFill method to fill the contour and get the accurate foreground target. 2. In this paper, we propose a cell-based anomaly behavior detection algorithm. Because of the low probability of abnormal behavior, short duration and no periodicity, it is difficult to model abnormal behavior. The strategy adopted in this paper is: firstly extracting the foreground of the training set and the test set, then dividing these prospects into meshes, extracting local features for each cell, and finally constructing the normal behavior model based on the features of the training set. Then, the characteristics of each cell in the test set are compared with the normal behavior model. If it is close, it is judged as normal behavior, otherwise it is abnormal behavior. Finally, fourteen video sets with different exceptions are established and tested. The experimental results show that the method proposed in this paper can effectively identify and track the abnormal behavior, and the algorithm can keep the frame rate around 11fps when the anomaly is detected, which can basically satisfy the daily monitoring work.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】