Biomedical Engineering Reference
In-Depth Information
being part of the foreground. In addition, we define the length of each edge on the
current RAG as wðE ij Þ¼ 0
5 ðd FG ðB i Þþd FG ðB j ÞÞ . Then, the graph growing method
:
works as follows:
(a) Find the skeleton of legs
The skeleton is defined as the shortest paths that connect the central nodes of
upper legs and lower legs. Here, the central node of upper legs is the closest
node to the centroid of upper leg blobs that is computed as the centroid of all the
blobs in the list of the candidate upper legs C UPLEG . Similarly, we can find the
central node of the lower right/left leg using blobs in C LRLEG or C LLLEG ,
respectively. After that, the shortest paths between those central nodes are
found using the Dijkstra's algorithm [ 11 ]. Note that only two central nodes
are available in a lateral view; thus, we only need to search for one shortest path.
The nodes along the shortest paths are called skeleton nodes.
(b) Perform graph growing from the skeleton
Select all skeleton nodes that satisfy the same geometric constraints (Sect. 3.2 )
as seed nodes, and add them to their corresponding list C Li . Then, starting from
a seed node, grow the graph to find all of the connected nodes that are in the
same body part list Li . Repeat this process for all of the seed nodes.
(c) Find the complete leg parts
Sort all of the nodes found in step b) that belong to the same body part Li in
ascending order of their distance to their corresponding skeleton node. Select
P
k
1 a j bA t 1
the first k ¼ min
k
nodes as the nodes of Li at time t , where A t 1
Li
Li
is the total area of Li at time t 1, and b is a factor that controls the rate of the
maximum allowed body size change during jumping process. The leg blobs are
directly obtained from the nodes found above.
As a special case, the head is tracked in a lateral view by finding the closest
connected nodes that have been initially labeled as {HEAD} from the head
centroid in the previous frame.
Figure 2 shows an example of initial association and semantic graph grow-
ing. Since our method expands graph nodes in an inward-out way from the
skeleton with an area constraint, it can exclude some nearby incorrectly
assigned nodes, like the bottom one shown in the third column of Fig. 2 .
3.4 Update Stick Figure
We merge the blobs found for each body part and use principal component analysis
method [ 12 ] to find its centroid, major direction, major, and minor lengths. Then,
compute new end points to update stick parameters accordingly. In our implemen-
tation, the upper legs in an anterior view are extracted together, which is further
divided into two parts by assigning pixels to their closest upper leg sticks.
Search WWH ::




Custom Search