Waypoints

one of the qualcomm track's prompts was building something that could only exist at the edge, no cloud needed, with the two brains on the board splitting the thinking and the reacting. that felt like a good fit for something safety related, where you genuinely can't afford to wait on a network round trip. falls are also just a real, common problem for people living alone, and most solutions either need someone to already be watching a camera feed or require the person to consciously call for help after the fact. i wanted something that reacts on its own the moment it happens, and that actually remembers where it happened so it can look out for you next time, all without sending anything off the device.

waypoints is a pocketable device built on the uno q. an onboard motion sensor watches for the signature of a fall, a moment of near weightlessness followed by a sudden impact. when that happens, it speaks a prompt through a small speaker and gives you twenty seconds to press a button if you're fine. if you don't, it treats the event as a real fall, and it remembers which zone you were in when it happened using a handful of bluetooth lighthouses placed around the space. the next time you wander back into a zone with a recent fall, it gives you a gentle spoken heads up to be careful there for a while. more lighthouses means more accurate zone coverage. everything, the fall detection, the audio, the zone tracking, and a small classifier that gets more confident over time, runs entirely on the device itself.

getting the full loop working end to end in an extremely short window, from raw accelerometer data all the way to a spoken warning tied to a remembered location, without ever touching the cloud. i'm also proud of how the two sides of the board ended up splitting responsibility. the microcontroller owns anything that has to be immediate and can't wait, the fall backstop, the countdown, the button. the linux side owns anything that benefits from more thinking time, the classifier and the lighthouse tracking. that split wasn't just architectural flavor, it's the actual reason the device stays trustworthy even if the smarter side is busy or slow.

a lot of this project was learning where inertial sensors quietly stop being trustworthy. a single imu can tell you a lot about orientation and impact, but the moment you try to turn that into actual position by integrating acceleration, small errors balloon fast, which is part of why i moved away from trying to track exact coordinates and leaned on proximity to fixed lighthouses instead. i also learned a lot about working with hardware that's still brand new, some things i expected to just work, like softwareserial, aren't supported yet on this board's core, so i had to learn to test each piece of the system in isolation before trusting the whole chain. and on the physical side, i got a much better appreciation for why shared ground matters the moment you start mixing power sources, a battery driving your speaker and a microcontroller driving your logic need to agree on zero volts or nothing talks to each other reliably.

as with any project idea realized over 36 hours, there's plenty of ways i could improve this! currently, the escort warns you every time you re-enter a fall zone. that might not be necessary for all individuals, in some cases, if someone hasn't fallen down somewhere in ages, that place could be considered easily maneuverable for them for a bit. for instance, if one doesn't fall near the kitchen in over a month, warnings for that area could be dismissed, unless they fall there again, of course.

ideally, that period of time could be further customized by the individual to their comfort. maybe it's better to be wary of those steep stairs for a while!

Built With

• arduinounoq