100 cars for what time period? One imagines that a much larger sample
(...) would be desirable. I wonder how hard it would be to assemble
more tracklogs like Allan's.
Allan, would you consider opening access to your collected points?
Joshua
The Netflix Prize is, I believe, $1m: http://netflixprize.com/
On Dec 1, 2006, at 9:13 AM, [EMAIL PROTECTED]
wrote:
I was hoping someone would ask this....
First of all, I think that aside from fun, this is an incredibly
valuable topic. This is a chance for your work to be a part of
saving many, many lives.
As for data- The DOT has a data set of 100 cars for a year from
northern Virginia (PVT data every few seconds). My proposal to
them has been to release a geographic third of the data to the
public for hacking and attempts to answer the questions I posed.
Then from the valid responses, they provide another third, with the
deliverable being a new map. Provider of the best map would
somehow move forward in support of a national effort in this
direction. The DOT would like some indication of interest in the
approach (hence this thread).
My personal belief is that the best map might come from a group
such as this, one without a significant stake in the current
approaches to mapping, and a strong geo-statistical background.
What sort of 'reward' would motivate a serious effort? I'm not
sure what the Netflix competition was.
-=Chris
[EMAIL PROTECTED] wrote: -----
To: [email protected]
From: joshua <[EMAIL PROTECTED]>
Sent by: [EMAIL PROTECTED]
Date: 11/30/2006 04:44PM
Subject: Re: [Geowanking] Probe based mapping of road network
This feels like a clustering problem, to me.
Is there any data that you have to share? Perhaps a netflix-style
competition might be in order.
Couldn't one analyze this on a per-region segmented datasets? (grid
partitions or whatever)
You can also probably identify time-of-day and day-of-week traffic
flow rates, and so on.
This project seems like an enormous amount of fun.
[EMAIL PROTECTED] wrote:
After watching this group for a while, thought it would be
interesting to bring up a topic I have been working on for several
years and see if I can get any help from the geowanking crowd.
Goal: Create highly accurate and complete digital maps of the
transportation network suitable for safety of life applications
with accuracy commensurate with future GNSS systems (decimeters).
It seems to me that this can only be done through a statistical,
probe based, approach since imagery and 'mobile mapping'
approaches are error prone with low revisit rates.
Problem : Given a very large set of vehicle PVT (position,
velocity, time) information,
1) derive the location of the centerline of every lane, along with
lane attributes such as direction and ability to cross to the
adjacent lane,
2) derive the location of all turn restrictions and traffic
controls, and
3) relate the PVT accuracy of the data to the accuracy of the
resulting 'map' for different quantities of data.
For extra credit, identify movements within lanes that indicate a
vehicle intends to turn, stop, or execute some other maneuver. Of
course, all of these answers must come with a statistical accuracy
metric.
Background : There are a lot of GPS units in a lot of cars
collecting a lot of data on where the cars (roads) are and how
they move (controls such as yields and stops). This data is then
thrown away. If this data can be captured (and there are efforts
underway to do this), how does one build a map of the roads and
all of the signs and signals that control the motion of vehicles?
I believe that the entire infrastructure that influences the
behavior of vehicles is captured in this data, and that, by the
central limit theorem, the data has ever increasing (and
quantifiable!) accuracy. This is exactly what is needed for map
based transportation safety systems currently under development.
This is one very promising way to address the 40,000+ fatalities/
$200B a year caused by accidents on US roads.
We spent a couple of years looking at this and devised a k-means
approach bundling data across the direction of travel to pull out
the lanes. The data could then be grouped by lanes to derive
centerlines. Stop signs and traffic lights were easy, we never
got to yields or speed limits. Our approach was successful, but
computationally intensive, and required that one work with the
entire data set rather than a Kalman filter approach where data
can be incrementally added to improve the solutions validity (or
indicate that the world has changed). We also did not get far on
the accuracy metrics. The key to this problem seems to be
grouping vehicles into like groups going from 'A' to 'B', where
'A' and 'B' are any two arbitrary points on the road network with
an accuracy of around 30 cm. We can 'generally' assume that a
vehicle is within 30cm of the 'lane center'. One problem, of
course, is that the accuracy of any individual vehicle's position
is generally somewhat larger than the lane width.
Does anyone know anybody working this (or similar) problems?
Any ideas on how to approach this from the geo-statistical crowd
out there? We came at this from an AI perspective, and I think a
geo-statistical approach might have gone a different direction.
Other thoughts?
-=Chris
PS- This approach is really promising for getting public, low
cost, accurate maps of transportation networks, and yes, there are
some serious privacy issues to work through. There will never be
unique identifiers in the data, and we can cut out the first and
last mile.
[EMAIL PROTECTED]
650/845-2579
_______________________________________________
Geowanking mailing list
[email protected]
http://lists.burri.to/mailman/listinfo/geowanking
_______________________________________________
Geowanking mailing list
[email protected]
http://lists.burri.to/mailman/listinfo/geowanking
_______________________________________________
Geowanking mailing list
[email protected]
http://lists.burri.to/mailman/listinfo/geowanking
_______________________________________________
Geowanking mailing list
[email protected]
http://lists.burri.to/mailman/listinfo/geowanking
