Machine Learning
Everyday Encounters: Camio Surveillance Cameras
May 11, 2017 Fred Schwaner

Internet Protocol Cameras, or IP Cams, have been around for 20 years. But they really only became affordable to the average consumer during the past 10 years, with the advent of cheap, low-to-middling quality cameras and affordable high-quality routers. For under $100, some tinkering and a large helping of patience, consumers could have an internet-connected camera transmitting video as well as audio. Most units had motion sensing, rudimentary low-light/night vision, and could record to a memory card to an IP address. Some allowed users to pan and tilt the camera remotely, and even to transmit audio back to the camera - to tell an intruder to scram or sing baby a lullaby.


But the reality never equal to the promise, mostly because of the low-quality of their cheap affordable lenses and, more than that, the relentless difficulty of false-positives from the motion sensor. IP Cams were, for many of us, our first real experience of attempting to filter live data to render it useful. Too much sensitivity on the motion sensor would trigger the camera ever minute or two as the light shifted through the curtains or leaves moved in the trees, while too little would render the camera useless. And there's always the chance of missing a critical moment to motion sensor lag or failure.


Recording and looping the entire stream is the safest solution from a security perspective. But even then, finding a key segment in a 24-hour stream is a challenge unto itself. And, for most budget users, the sheer volume of data was enough to fry most SD or MicroSD cards after a short while. Until Camio, the solution has been to spend on better and better cameras.


Camio's approach is to rent an intelligence layer to its users’ cameras- whether they be connected IP Cams or decommissioned smartphones. For a monthly fee, Camio users can store 30 days’ worth of continuous video in the cloud, receive digests and alerts of notable events or anomalies, and use its smart search function to quickly find events without having to watch the whole stream. The camera’s feed becomes a dataset against which Camio runs its own constantly evolving rules and algorithms. The idea is that by adding modern analytics and machine learning to old cameras, the original promise of useful video surveillance can be delivered without the need for a hardware upgrade.


The potential is exciting. Machine learning allows users to effectively customize each camera for its specific use and location, and to cut through the noise to deliver only relevant alerts, digests and information.


That's today. But the same algorithm with higher video resolution and bandwidth could perhaps enable the system to identify individuals and things more specifically - it's just a matter of degree. Mom just got home or It's 4 o’clock and Jimmy isn’t back from school yet” or “A car just pulled up, license plate ABCDE1, parked for under 4 minutes and then drove off” or “The UPS truck just pulled up.” And, that could connect that with the August Smart Lock, so that the door opens for family members, for some friends at certain times, and so on.


It’s still very early days, but the addition of cloud-enabled intelligence to the Internet of Things holds vast promise for many of those old things and their users.

See more examples of Machine Learning in our Everyday Encounters blog series >>

Fred Schwaner
Machine Learning Engineer