Is my living room occupied ?
Accurate — say, 0,1% error rate — detection of room occupancy is a graal in Home Automation systems. We have now nice, low cost, pretty accurate detectors that use infrared grids to detect movement. That solves mainly the issue for rooms such as kitchens, halls and — up to a point — cellars and attics. The strategy fails miserably for living and bedrooms, where people tend to lie at rest;
A few other strategies are promising : Use still infrared detectors (expensive). Use beam rays and count people entering and leaving rooms (young technology for home, fails the waf test, fragile?).
The most promising technology seems to be object classification, a well researched field of AI, pretty much solved recently with CNNs (with special mention for the YOLO vX line of neural nets)
So, given a 5$ detector and a 25$ camera, the idea is: whenever there is a movement or end of movement in the room; check with the camera, feed to the classifier, and count whether there is any person. Problem solved!
Problem solved ?
Well, not really…
In a living room — and I guess even in a bedroom — people tend to lie and behave differently than in your standard street.. They don’t move. they lie on couch in unusual position. Light is dim while tv is playing. They are naked (well I cetaintly am, sometimes) or covered with plaid. Standard training datasets dont include such cases and neutral nets quiclky misclassifies.
One road to solve the issue is to build a new training set, perhaps with knowledge transfer.
Another road, the one I want to purse, is to exploit the geometry and constraints of the room. People odnt lie anywhere. they enter by a door (CNN pretty accurate here) then move till they reach a still region — a couch , a table — where we know that the error rate of detectors skyrockets. Yet, if a follow up of people trajectory shows that they entered but didnt leave the room; we are pretty confident that they are still there, somewhere, most likely near the place the Net last saw them…
Which leads us to Bayesian inference..
