基于自适应聚类算法的小群体检测与跟踪

发布时间：2018-05-17 14:55

本文选题：小群体 + 数据关联　；参考：《北京交通大学》2017年硕士论文

【摘要】：小群体检测与跟踪是智能视频监控系统的关键技术,也是异常事件检测、行为理解、场景理解等更高层次的视觉任务的基础。小群体指的是在接近的运动区域中,若干具有动作一致性的人群。监控视频中的小群体不仅反映了社会行为与安全问题,具有广泛的应用,而且是计算机视觉领域非常具有挑战性的问题。小群体的检测和跟踪依赖于对个体运动目标的检测与跟踪,同时也依赖于群体特征的描述和群体获取算法。这涉及到图像处理、模式识别和机器学习等多个领域的知识,因此具有较高的理论研究价值。近年来,许多群体分析的算法不断的被提出,但是由于监控视频中人群运动的活动分布状态、小群体的结构动态变化频繁、相互遮挡问题以及复杂背景问题等都给小群体分析带来了不小的挑战。为此,本文研究了一种基于自适应聚类小群体检测与跟踪的方法,该方法通过改进多目标跟踪,优化轨迹的相似度测量,并进行自适应聚类,来发现小群体并对小群体跟踪。本文的主要工作分如下两点:(1)对运动目标个体,提出了基于双向速度的预判轨迹拟合的多跟踪算法。它是针对监控视频中人群可能出现相互遮挡、复杂背景等问题提出的,目的是为后继的基于自适应聚类的小群体检测算法提供更准确的输入。在公共数据集FM_dataset和SMOT数据集上实验结果证明,提出的基于双向速度的预判轨迹拟合的方法能够在目标短时遮挡时恢复目标的时空运动轨迹,大大提高了对多目标跟踪的准确性。(2)提出了基于自适应聚类算法的小群体检测与跟踪算法。它是根据多目标检测与跟踪获得的时空轨迹,通过进行轨迹相似度的度量,对度量场景的直方图统计分布,自适应的确定分割小群体的阈值,进而通过聚类算法自适应的判断小群体的分裂与合并。最后基于正确率、缺失率、误判率等评价指标,在FM_dataset数据集上验证了本文提出方法的鲁棒性。
[Abstract]:Small group detection and tracking is the key technology of intelligent video surveillance system. It is also the basis of higher level visual tasks such as abnormal event detection, behavior understanding, scene understanding and so on. A small group refers to a group of people who have a consistent movement in a close movement area. The small groups in surveillance video not only reflect the social behavior and security problems, but also are very challenging in the field of computer vision. The detection and tracking of small groups depend on the detection and tracking of individual moving targets, as well as on the description of population characteristics and the algorithm of group acquisition. It involves many fields such as image processing, pattern recognition and machine learning, so it has high theoretical value. In recent years, many group analysis algorithms have been proposed, but because of the distribution of crowd movement in the surveillance video, the structure of small groups changes frequently. Mutual occlusion and complex background problems all pose a great challenge to the analysis of small groups. In this paper, a method based on adaptive clustering for small population detection and tracking is studied. This method can find small groups and track small groups by improving multi-target tracking, optimizing similarity measurement of trajectory, and carrying out adaptive clustering. The main work of this paper is as follows: 1) A multi-tracking algorithm based on bidirectional velocity predictive trajectory fitting is proposed for moving target individuals. It aims to provide more accurate input for the following small group detection algorithm based on adaptive clustering. The experimental results on the common data sets FM_dataset and SMOT show that the proposed method based on bidirectional velocity predictive trajectory fitting can restore the temporal and spatial trajectory of the target in a short period of time occlusion. The accuracy of multi-target tracking is greatly improved. (2) A small group detection and tracking algorithm based on adaptive clustering algorithm is proposed. It is based on multi-target detection and tracking of the space-time trajectory, through the measurement of trajectory similarity, the histogram statistical distribution of the measurement scene, adaptive determination of the threshold for segmentation of small populations. Then the clustering algorithm adaptively determines the split and merge of small population. Finally, the robustness of the proposed method is verified on the FM_dataset dataset based on the correct rate, missing rate, misjudgment rate and so on.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【参考文献】