We have now completed the prestudy of our CPS project. Apart from project stuff like funding, resources and agreements, we made some interesting findings already.
1) Lookup algorithms.
If we know where our wifis are exactly, we have claimed in Indoor CPS that we are below 10m median error. This is ofcourse depending on density of wifis. We have now made a test with a corridor with 4 wifis 20m apart. Our current CPS indoor service gives a mean error of about 4m and median error of about 3m. We have tested several other algoritms from research articles like MLE and RANSAC variants, but none of them improve the error. The only way to improve the error is to add knowledge about last position(s), thus use track information. Then we can add Kalman or Particle filtering and other algorithms and get the mean and median error down to less than half.
Our conclusion: Current CPS indoor lookup algorithms works well. We should add track lookup or history to improve further.Thus if we know the exact position of wifis, we should be able to get accurate indoor location.
2) Learning algorithms (which is really main focus for our project).
We hate to manually insert our indoor wifis into our database. Even if we have our Location-API app as tool and can edit them, it is a lot of manual work to get exact correct location of each wifi into our database. Then how can we instead automatically calculate the location from our measurements?
We have in the prestudy developed a simple process to get data from our database as well as our Location-API app into Matlab. In Matlab we can much more easily evaluate different locations and ideas. In the pictures above, we show some tracks (we walked only the isles, not into any stores) made for a shopping center where we have not made any manual learning. Blue dots are the track positions where we estimate the scan was made and red circles are showing where we think the wifis are. To the left is current CPS API performance. Since we base our database on crowd sourcing of data with GPS, we do not get far into the building since there is no GPS available, thus most wifis are close to the entrances. But to the right we have made a first run in our algorithms. Now the wifis (red circles) have moved into the building and the scan track with it. We are not there yet and there are much more work to do, but initial results are encouraging!
As you can see we have started in 2D and now the fun starts to improve our algorithms! Stay tuned!
Recent Posts
