I don’t know if there is some sort of consensus. I have benne looking at different strategies to get the best presence detection. My conclusion so far is, that there doesn’t exist one particular method that will cover everything. All methods have it’s own strengths and weaknesses. So I basiclly believe that a good solution is to combine different methologies to get the optimum solution.
I see the sensors grouped in personal / non personal sensors and portable / non portabel sensors.
I guess almost all non personal sensors (at least all those I can think of), is also non-portabel:
- Door/Window contact
- PIR Sensor
- Proximity sensor
Where I have knowledge of both kinds of Personal sensors:
- OwnTracks (Portable)
- Beacon (both)
- Wifi (Non Portable)
- Bluetooth (Non Portable)
Amongst others…
The non portable sensors has the strength that “false positives” is very close to none. That is since it typically detects if a eg. Wifi newtwork is visible, and thus that defines the geofence. On the other hand It doesn’t deal very well with “false away”. (E.g., if I turn off my phone while at home, presence will be offline)
The portable sensors is OwnTracks or similar. OwnTracks, is event based, so where it really doesn’t matter if I turn off my phone at night (since the transition wouldn’t occur). I considder it as quite stable and accurate, Though I sometimes (rarely) see that the transition doesn’t happend.
The nonpersonalized sensors is also interesting, but I believe that they should be more an supplement, than a primary decission.The Wasp In A Box Algortihm has some thoughts about this. Seen from my perspective, the usage of those sensors could not be of much use alone, but I you build a model and monitors a series off events (eg. pir sensors and door sensors), you could have quite a good guess if somebody is leaving or comming. The problem is that you need to build an individual model, so the precision of the outcome is highly related to how good the model is.
The more I think about it, I believe that I could obtain more or less the same precission by combining the two personalized sensor types. At least if I would create a genric algorithm that is usefull for every body. Maybee some selftraining netwok could be usefull to build a good model.
I still believe that there is a lot off information in the non personalized sensors.
Actually I think, that if I setup a couple of Raspberry Pi’s with a WiFi dongle in monitor mode and also scans for particular Bluetooth MAC’s, combined with OwnTracks. I would have a hit rate close to 100% - at least in my case.