LoRa Security Camera Ignores Non-Threats Using Neural Network

What do you do when you need a security camera for a remote building that isn’t quite in WiFi range? With its impressive range specs, but very limited data bandwidth, LoRa wouldn’t be the traditional choice for such an application, but Tegwyn Twmffat has figured out how to make this transfer method work.

The trick here is that instead of transmitting a video stream, a la pretty much any other type of camera setup, it captures images and processes them locally to determine the threat level. This is then passed to a base station that beeps in accordance with how close the intruder gets to the sensor.

It’s an easy enough concept to grasp, but actually getting the system running reads like a who’s who (what’s what?) of maker technology. A Raspberry Pi and camera to actually take the image, then passes it along to an Intel Movidius Neural Compute Stick. The Stick then classifies the approaching object into interesting: e.g human, or not a big deal: cow, chair, sheep, sofa, etc. If it is interesting/threatening the LoRa connection simply transmits a small bit of data to note there’s something lurking about. This is picked up on the other end by an Arduino MKR WAN 1300, which beeps and flashes in accordance with detection distance, inferred by the size of the intruder detection box.

As seen in the video above, it’s a clever configuration, and successfully ignores an generally benign dog, while beeping away when a human intruder — Twmffat, that is — comes into frame.