Abstract:

A novel unsupervised method based on motion pattern analysis is proposed to detect abnormal activity in video flow. In order to effectively describe the object information in the video scene, spatial-temporal descriptors are extracted for each foreground pixel combined with block region information, and the video is represented by visual words which are generated through the bag-of-words model. Then, the sparse topic model is proposed to obtain the motion topics in video. Through this model, the motion patterns can be discovered and used to encode each video document. Finally, through the analysis of the reconstruction accuracy and the motion pattern composition, the test video clip can be determined whether it contains abnormal activity or not. Experiments are conducted on QMUL Junction dataset. The results demonstrate the superior efficiency of the proposed method.

Key words: motion pattern, spatial-temporal descriptors, visual words, sparse topic model, reconstruction accuracy

摘要： 提出了一种基于运动模式分析的无监督方法用于对视频中的异常行为进行检测。为了有效描述视频场景中的各种目标信息，对每个前景像素点提取了时空描述符，再结合块区域信息并通过词袋模型生成视觉单词，对视频进行表示。提出了一种稀疏主题模型，用以获取视频中的运动主题。通过该模型可以查找出视频中典型的运动模式，并可以此对各视频文件进行编码。通过分析重构精度和运动模式组成对测试视频文件进行检测，判断其中是否包含异常行为。实验在QMUL数据集上进行，实验结果证明了所提方法的有效性。

关键词: 运动模式, 时空描述符, 视觉单词, 稀疏主题模型, 重构精度