基于动态粒子流场的视频异常行为自动识别

Dynamic particle flow field based automatic recognition of video abnormal behavior

为了有效实现视频异常行为的自动识别,基于动态粒子流场,将视频运动对象的运动行为,映射为有效反映其运动变化状态的动态粒子流,同时提取度量不同场景内容下的运动方式各异的异常行为的显著性运动特征,进行异常行为的分类与识别。对来自不同场景并具有不同运动行为方式的公开视频测试序列的实验表明,本文方法无需跟踪运动对象,也无需预先采集异常行为样本进行学习与训练,可在多种条件下实现视频运动对象异常行为的有效自动识别。

How to efficiently realize automatic recognition of abnormal behavior for intelligent video surveillance is a key problem. A method has been developed that the dynamic particle flow field from video is got based on the Lagrange dynamic system equation and self-adaptive determination of time interval in the equation. Some motion behaviors for motion objects in video are mapped to the dynamic particle flows which can be used to describe their motion variation states. Some significant motion features for abnormal behavior with different motion styles from different scenes have been extracted to classify and recognize the abnormal behaviors. Some open video test sequences from different scenarios with different behavioral patterns are selected to perform experimental verifications and comparisons. Experimental results show that abnormal behavior can be automatically recognized efficiently in various conditions where it is not necessary to track motion object or collect abnormal behavior sample in advance for learning and training.