Algorithm 1. The online segmentation process

Initial: Start one segment from the frame $start$, let $F_{start}$ and denote the feature vector of the frame and the sign of $E_d(start)$, respectively. For each feature vector $F_i$ (The subscript $i$ indicate the frame index) Step 1: Compute the potential difference $E_d(i)$ and its sign $S_i$. Step 2: If $S_i==S_{start}$ If $i-start>L{P}_{\max }\&S_{start}==0$ /*$L{P}_{\max }$ is the maximal length of pose feature segments. */ Complete the segment at the frame i–1; /* Pose feature segments. */ Else Continue; End Else /* $S_i\ne S_{start}$ */ If $i-start<L{M}_{\min }\&S_{start}\ne 0$ /*$L{M}_{\min }$ is the minimal length of motion feature segments. */ $S_{start}=S_j$; /* Eliminate tiny motion feature segments. */ Continue; Else Complete the segment at the frame $i-1$; /* The segment type is decided by $S_{start}$. */ If $S_i\ast S_{start}==-1$ /* Two adjacent motion feature segments */ Insert one pose feature segment which is only composed of the features $F_{i-1}$ and $F_i$; End End End Step 3: Start one new segment from the frame i, let $F_{start}=F_j$ and $S_{start}=S_j$; End