多尺度格兰杰因果在行为分析中的应用

Application of multiscale Granger causality in activity analysis

考虑到复杂场景中的行为具有分级结构的特点,文中提出了一种基于多尺度格兰杰因果的行为分析算法框架。首先提取视频底层运动特征,并进行量化;然后基于概率主题模型对场景中的行为进行检测,实现多变量时间序列的提取;进一步基于状态空间模型对多变量时间序列进行表示;最后利用状态空间模型参数,实现行为间的格兰杰因果测量;通过实际的视频监控数据对所提算法框架的有效性和可行性进行了验证。实验结果表明:与固定尺度的行为分析算法相比,文中所提出的框架能够更有效地检测出行为间不同尺度的依赖关系。

Considering the hierarchical structure of activities in complex scenes,a framework of activity analysis based on the multiscale Granger causality is proposed.First the low level motion features are extracted and quantized.Then activities are detected by probabilistic topic model and furtherly converted into the multivariate time series.The multivariate time series is further fitted based on the state space model.Finally,the state space model parameters are used to realize the Granger causality between the activities.The effectiveness and feasibility of the proposed method are verified by real video surveillance data.The experimental results show that the proposed framework can detect the dependences between activities at different scales more effectively.