"Average paths"
Hi guys!
Suppose I have data showing how n people found a way from A to B through an obstacle course (open area with obstacles).
I now want to filter this data and visualize some sort of "average paths" for analyzers. The point is that the operator wants to be able to look at this modified view and say something like "ok, it looks like most people passed the rock, and continued straight to the goal. Some people chose to stop by the bar, but almost noone ever came close to the forest ".
I was just wondering if anyone has done anything similar and can share some ideas on how to condense a really large number of paths into a more intuitive representation and visualize this.
This does not need to be very exact, it's just going to be used as a "overview".
Any ideas greatly appriciated!
Regards
/Alex
[855 byte] By [
ArneWeisea] at [2007-11-26 17:48:28]
# 1
The first thing which springs to mind is use of alpha (transparency). With suitable normalisation, you could arrange for the most heavily used pixel to be solid or near-solid; if you want to convey information on how long people wait then using multiple channels - blue for moving; red for waiting, with alpha indicating total time waited summed over all people.
# 2
> The first thing which springs to mind is use of alpha
> (transparency). With suitable normalisation, you
> could arrange for the most heavily used pixel to be
> solid or near-solid; if you want to convey
> information on how long people wait then using
> multiple channels - blue for moving; red for waiting,
> with alpha indicating total time waited summed over
> all people.
Seems like a nice idea. There is still one problem though, I would like to show direction in the movements aswell. Maybe you could superpose paths onto your proposed colored visualization?
When reading up on this, I figure that I would need to use some sort of time-series smoothing and then apply clustering of trajectories with automatic cluster selection. The MSE of the euclidian distance between trajectories along with some sort of direction indication could serve as a good clustering criteria.
Any good algorithms for this?
# 3
In case anyone is interested i used dynamic time warping (with boundaries) and/or least common subsequence trajectory matching with k-means clustering for this. Also needed to incorporate velocity measures and resampled the trajectories to a predifined number of samples per distance unit to compare them. It's great fun to play with the tool, but has high time complexity at the moment. But all in all, works great! :)
