Now everyone is back in office eager and inspired to make our research progress even further!
But our research has continued during summer despite the nice weather and we are now below 10m in median scan position error for our indoor reference track at the Nova shopping center in Lund.
1) Learning algorithms (main focus for our project).
To repeat, our goal is no manual learning and no calibration! All learning should be based on our crowd sourced data. We have now added track history, thus we detect a user track with scans at regular intervals, like every 3 or 5 seconds, and use that for improving our bundler optimization. The improvement is about 10%. We will not rely on this, but this shows that for a venue where better accuracy is required: Just use our app and walk around as much as possible and walk one time around the building on the outside.
Another big improvement we have made, is to allow damping factor to be optimized for each wifi. Thus taking into account the difference of a wifi perfectly placed in the mall corridor and a wifi way back behind a wall in the store.
Our improvement is shown in figure above. Blue dots are the track positions where we estimate the scan was made and red circles are showing where we think the wifis are. Figure left is our ground truth track. Center figure is current CPS API track. Right is our latest Indoor CPS track based on new algorithms. Errors are now about one third of initial CPS errors for this venue (9.5m median scan position error and 12m median wifi position error).
Going forward we will now make this optimization more robust for larger areas and many buildings and see how we can optimize algorithms for speed when using our enormous amount of data. 2) Lookup algorithms. We have started to implement improved lookup algorithms on our real servers. Will soon start running tests with old and new algoritm for comparison. Still as before, no major improvements are expected from better lookup algorithms. The improvement comes from a much better quality database based on the better wifi position accuracy from our improved learning. 3) Our tool, the Location-API app, can now show track on a map based on current CPS algorithms. Download and check yourselves. The position is unfiltered and jumps around a little, so we plan to implement a Kalman filter in our Location-API app when we have time to make the tracks there nicer.
But to get really good indoor CPS with our app, we first need to recalculate the relevant area (tiles) in our database. Still more work required before it is time to update our database. Stay tuned!
Recent Posts
