机场智能视频监控中异常行为检测与目标跟踪算法研究

发布时间：2018-06-02 16:35

本文选题：智能视频监控 + HOFO　；参考：《南京航空航天大学》2017年硕士论文

【摘要】：随着中国民航业的迅猛发展,机场所面临的安全压力日益增大。本文针对传统视频监控系统局限于人力监视缺乏主动检测识别异常事件的能力,对智能视频监控系统(Intelligent Video Surveillance,简称IVS)中的异常行为检测算法与目标跟踪算法进行了研究,旨在为高性能的智能视频监控系统的开发与实现提供思路和参考。本文主要研究内容如下:针对异常行为的检测问题,本文对传统的基于光流方向直方图(Histogram of Optical Flow Orientation,简称HOFO)的异常检测方法进行了改进。传统的光流方向直方图的计算仅限于对光流方向的简单统计,对图像信息的描述存在不足,为了提高对图像信息的表达能力,本文中对直方图的统计方式进行了改进,将光流能量加权到直方图的计算中,提出了一种基于加权光流能量的HOFO特征的异常行为检测算法。实验表明,改进后的算法与原始算法相比检测准确率得到了一定程度的提高。针对上述异常检测算法检测速度与准确率低的问题,本文将基于卷积神经网络(Convolutional Neural Networks,简称CNN)应用于异常行为的检测,提出了一种基于卷积神经网络的异常行为检测算法。该算法不需要设计特征提取器,可以直接将图像作为输入,同时又采用了局部感知和权值共享的方法,大大加快了算法速度。实验表明,该算法相对于上述异常行为检测算法不仅加快了算法速度而且提高了检测准确率。针对现有跟踪算法遇到遮挡、形变、以及光照变化而引起的跟踪失败问题,本文提出了一种融合表观特征与深度特征的目标跟踪算法。首先用大量行人数据库对CNN网络进行训练,然后用训练好的CNN网络提取目标区域的深度特征,同时计算目标区域在HSV空间的颜色直方图,将深度特征与颜色特征进行联合得到整体特征。最后在粒子滤波框架下对多个假设状态进行估计,获得目标的最优位置,得到跟踪结果,并进行模板更新,最后根据粒子的退化情况,进行重采样。实验表明,本文跟踪算法获得了良好的跟踪鲁棒性。最后,设计了异常行为检测与目标跟踪系统并在Matlab平台上进行了仿真实现,验证了本论文所研究算法的有效性和实用性。
[Abstract]:With the rapid development of China's civil aviation industry, the airport is facing increasing safety pressure. In this paper, the traditional video surveillance system is limited to human monitoring, which lacks the ability to detect and identify abnormal events. In this paper, the detection algorithm of abnormal behavior and the algorithm of target tracking in intelligent video surveillance system (Intelligent Video Survey) are studied. The aim is to provide ideas and references for the development and implementation of intelligent video surveillance system with high performance. The main contents of this paper are as follows: aiming at the problem of abnormal behavior detection, this paper improves the traditional anomaly detection method based on histogram (HOFOO). The traditional calculation of optical flow direction histogram is limited to the simple statistics of the optical flow direction, but the description of the image information is insufficient. In order to improve the expression ability of the image information, the statistical method of the histogram is improved in this paper. In this paper, the optical flow energy is weighted to the histogram, and an anomaly detection algorithm based on the weighted optical flow energy HOFO feature is proposed. Experiments show that the detection accuracy of the improved algorithm is improved to some extent compared with the original algorithm. In order to solve the problem of low detection speed and accuracy, this paper presents an anomaly detection algorithm based on convolution neural network (CNNs) based on Convolutional Neural Networks (CNNs). The algorithm does not need to design a feature extractor and can directly use the image as the input. At the same time, it uses the method of local perception and weight sharing, which greatly accelerates the speed of the algorithm. Experimental results show that the proposed algorithm not only speeds up the speed of the algorithm but also improves the detection accuracy compared with the above algorithms. In order to solve the problem of tracking failure caused by occlusion, deformation and illumination change, a target tracking algorithm combining apparent features and depth features is proposed in this paper. Firstly, the CNN network is trained with a large number of pedestrian databases, then the depth features of the target area are extracted by the trained CNN network, and the color histogram of the target area in the HSV space is calculated at the same time. The depth feature is combined with the color feature to get the whole feature. Finally, several hypothetical states are estimated in the framework of particle filter, the optimal location of the target is obtained, the tracking results are obtained, and the template is updated. Finally, according to the degradation of particles, the sample is re-sampled. Experiments show that the proposed tracking algorithm has good tracking robustness. Finally, the anomaly behavior detection and target tracking system is designed and implemented on Matlab platform, which verifies the validity and practicability of the algorithm studied in this paper.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN948.6

【参考文献】