智能空间下基于时序直方图的人体行为表示与理解

Human Activity Representation and Recognition Based on Temporal Order Histogram in Intelligent Space

为了实现在智能空间下理解人体行为,该文提出了一种基于时序直方图的人体行为表示方法,并设计了基于softmax回归分析的行为理解算法.将传感器的输出状态所对应的离散值的编号定义为事件项;提出了事件直方图的概念以表征事件项的发生频率,提出了时序直方图的概念以表示传感器序列的时序信息;为了降低过拟合的风险,设计了时序模式选择算法以得到候选特征集,基于互信息准则进行特征选择,完成行为的表示;最终,基于选择的特征子集及softmax回归分析实现了人体的行为理解.实验结果表明,该文提出的特征表示方法可以较好的表示外部传感器序列捕获的人体行为模式,同时softmax回归分析可以较好的实现人体行为的理解,并取得了良好的效果.

In order to recognize human activity in intelligent space, a novel activity representation method based on temporal order histogram was presented, and softmax regression was used to design the recognition algorithm. Ambient sensors were used to collect data and based on these data human activity recognition was realized. In order to model the sensor data sequences, the concept of event item was put forward. Event histogram and temporal order histogram were put forward to represent the human activity. In order to decrease the risk of over fitting, feature selection algorithm was designed to obtain the candidate temporal order feature sets, and mutual information was used to select the final features. At last, softmax regression was used to realize the human activity recognition. Experimental results showed that the method can better realize human activity recognition.