Following on from the GPS accuracy tests carried out on the roof of the OU Biology building, we’ve been doing a set of tests on the balcony of the Knowledge Media Institute (KMI). We’ve tried as far as possible to do a comparable set of tests. The KMI balcony is on the top floor of the building and faces south(ish), the wall of the building blocks part of the sky, but only part of it.
An upturned bookshelf is g-clamped to the railings of the balcony and nine positions (10 cm apart) are marked along the middle of it. The GPS devices can then be placed on any of the nine positions and are held in place either by an elastic shock cord or gaffer tape. The Android phones are put in ‘Aquapac’ waterproof bags which are also tied to the cords on the shelf in case they fall.
We repeated a set of tests on two consecutive days to see the effect of using having an internet connection. On both days we used five phones (three HTC Desires and two HTC Wildfires). The phones had a WiFi connection to a local wireless network and logged their GPS location once per minute to a local web server. On the first day the network was not connected to the internet, on the second day a 3G dongle (connected to the three network) was plugged into the WiFi router and provided internet access for all five phones.
The device positions and configurations were as follows:
- Position 4 – HTC Wildfire (id-6) using internal GPS
- Position 5 – HTC Wildfire (id-7) using internal GPS
- Position 6 – HTC Desire (id-5) using internal GPS
- Position 7 – HTC Desire (id-2) using internal GPS
- Position 9 – HTC Desire (id-4) using external GPS
The tests ran for about 3 hours 20 minutes on both days. Although the data patterns vary across the phones, the accuracy is comparable to the previous sets of tests. The phones’ GPS readings are generally within 2 or 3 meters of their actual position. The phones seem to lock onto a value for a while, so the plots look like they are stepped. The external GPS (a Nokia LD-3W) although comparable to the internal GPS for most of the time, had an unexpected spike on both days.
Here are the plots for the first day (no internet connection).
The following day we repeated the tests, but included internet access (i.e. a 3G mobile broadband dongle connection) over the local network.
The plots are surprisingly similar across the two days. For each phone the trends are generally comparable. It may be that the internet access used by the AGPS service is not so crucial. At the moment, we don’t know for sure if the phones on the first day (that did not have internet access) were able to use a previously cached AGPS dataset. If anyone knows for sure please let us know.
To conclude, more replication is required to see if the results hold. However, so far an accuracy of two to three meters seems like a reasonable expectation. One of the Wildfire phones (id-6) dropped a lot of data on both tests, it could be that it needs a judicious kicking, but we’ll first try checking for application upgrades and validating the settings (its already running the current version of Android).