This paper describes methods to create motion
graphs that can be used to create continuous motion with
high quality transitions automatically.
First it compares every pair of motion and finds local
minimas of where frames differe the least.
With a certain threshold, we disgard some that aren't good enough.
After connecting the graph we find the strongly connected
components and only use them so we don't end up in a dead end
The paper also suggests ways to reduce
graph search time by using a branch and bound strategy.
Improving the distance function between frames and
proper threshold management would be needed from data sizes
going out of hand. For constraints, I wonder how well we
can deal with footskates if leg stretching isn't allowed.
Even with branch and bound, maybe good quality graph search
can be too expensive for real-time crowd simluation.