视频中的稀疏多目标跟踪和轨迹异常检测研究

发布时间：2018-03-22 19:37

本文选题：前景检测　切入点：多目标跟踪　出处：《西南交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：随着意外事故、犯罪和恐怖活动的增加,公共安全显得越来越重要。面对这些突发事件,智能视频监控系统能够及时的给出预警信号或报警。与传统的人工监控摄像头相比,智能监控系统能够节省大量的人力、物力和财力,并且能够更加高效的对这些合法的视频监控数据实现自动或者半自动的解释和分析处理。在智能监控系统的研究中,视频前景检测、多目标跟踪和异常行为识别研究作为比较新的研究方向,已经成为计算机视觉和模式识别领域的研究热点,它们的研究对于提高智能监控系统的性能具有非常重要的意义。本文通过对视频前景检测、多目标跟踪和异常行为识别领域的算法分析,对智能监控系统的开发进行了深入的研究。主要完成以下几个方面的工作： 1.归纳总结了前景检测领域常用的运动目标检测方法,并对常用的运动前景检测方法进行介绍,提出一个改进的基于混合高斯模型的运动目标检测方法,大大提高了以往基于混合高斯模型的前景检测的鲁棒性和准确性,其抗干扰能力显著增强。 2.在跟踪阶段,针对单固定摄像头,提出一个稀疏的多目标跟踪系统框架。该框架重点是将单目标跟踪很好的TLD算法和关联矩阵结合起来,有效解决多目标跟踪过程的合并遮挡问题。在目标合并处理阶段,对合并的目标加窗且引入双三次插值算法对初始化的目标和所加窗口进行同比例超分辨缩放。该操作能很好地解决大目标的计算复杂度高和小目标的不能正常初始化问题。对于关联矩阵的一些特殊情况进行特殊处理。最后在滤波阶段,该框架用分数阶卡尔曼算法代替卡尔曼算法进行滤波,不仅能够降低机动目标的观测噪声,还能在间隔跟丢时准确地预测目标的位置。 3.在基于轨迹的异常检测阶段,本文提出一个基于时间分割的多特征表示的轨迹异常检测方法。首先提出一种新的轨迹特征表示方法,该方法由六个特征空间组成：1)轨迹的方向和长度,2)轨迹的平均位置,3)初始位置、轨迹分割片段的时间长度、分割片段的方向,4)分割片段序列的平均速度序列,5)分割片段序列的平均加速度序列,6)整条轨迹的最大加速度。接着利用监督型的支持向量机分类算法来对轨迹特征集进行训练、检测。该方法提高了轨迹异常行为的异常检测率和识别准确度,降低了虚警率。同时,由于不需要对训练和测试样本进行缩放处理从而大大提高了该方法的实用价值。
[Abstract]:With the increase of accidents, crime and terrorist activities, public safety becomes more and more important. In the face of these emergencies, intelligent video surveillance system can give early warning signal or alarm in time. Intelligent monitoring system can save a lot of manpower, material resources and financial resources, and can more efficiently interpret and analyze these legitimate video surveillance data automatically or semi-automatically. As a new research direction, video foreground detection, multi-target tracking and abnormal behavior recognition have become the research focus in the field of computer vision and pattern recognition. Their research is very important for improving the performance of intelligent monitoring system. Through the algorithm analysis of video foreground detection, multi-target tracking and abnormal behavior recognition, the development of intelligent monitoring system is deeply studied in this paper. 1. The common moving target detection methods in the field of foreground detection are summarized, and the commonly used motion foreground detection methods are introduced, and an improved moving target detection method based on mixed Gao Si model is proposed. The robustness and accuracy of foreground detection based on mixed Gao Si model are greatly improved, and its anti-jamming ability is greatly enhanced. 2. In the tracking phase, a sparse multi-target tracking system framework is proposed for a single fixed camera, which focuses on combining the TLD algorithm with the correlation matrix. Effectively solve the merge occlusion problem in the multi-target tracking process. The combined target is windowed and the bi-cubic interpolation algorithm is introduced to scale the initialized object and the added window in the same proportion. This operation can solve the problem of high computational complexity of large target and abnormal initial value of small target. To deal with some special cases of incidence matrix. Finally, in the filtering stage, The frame uses fractional order Kalman algorithm instead of Kalman algorithm to filter, which can not only reduce the observation noise of maneuvering target, but also predict the position of target accurately at interval and loss time. 3. In the phase of locus based anomaly detection, this paper presents a method of trajectory anomaly detection based on multi-feature representation based on time division. Firstly, a new trajectory feature representation method is proposed. The method consists of six feature spaces, the direction and length of the trajectory, the average position of the trajectory, the initial position and the time length of the segment. The direction of the segment is 4) the average velocity sequence of the segment sequence is 5) the average acceleration sequence of the segment sequence is 6) the maximum acceleration of the whole trajectory is obtained. Then the supervised support vector machine classification algorithm is used to train the trajectory feature set. The method improves the detection rate and recognition accuracy of trajectory anomaly behavior and reduces the false alarm rate. At the same time, the practical value of the method is greatly improved because it does not need to scale the training and test samples.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.6

【参考文献】