When you have kids, you know they leave doors open. In this case, the gate in our garden. This does not have to be a big problem, except when there are bikes unlocked in the backyard. After all: we don’t want them to be stolen! I don’t want to equip my bikes with sensors and I also don’t want to put sensors on this wooden gate in my fence.
I used a camera (Xiaomi Dafang in my case) for motion detection. Using some cloud technologies I process these images and publish these ‘facts’ as sensors in Home Assistant.
It’s a matter of teaching. I had to teach the model for different conditions. For example, when it rains the fence/gate gets much darker. Another example was the plants, because it’s spring, the backyard changes a lot which affected the detection of the door (grape grew before the barn door ) These kind of problems where quite easy to fix by some new example images/trainings. After that it’s pretty good I have to say!
Have you characterised the accuracy of classification for the different classes, and perhaps established a minimum number/range of images to get satisfactory accuracy?
Awesome! You and I took a similar approach on how to solve this, however I’m it all locally using TensorFlow/DarkFlow/YOLO.
Blue Iris is configured to send an MQTT message when any of my cameras detect motion, my script pulls down a JPEG of the camera that saw motion, classifies it in YOLO, then publishes the results via MQTT. It’ll announce people outside via Alexa by sending a string to an MQTT topic that the openHAB Alexa TTS binding watches. I can also have it classify each camera on a regular basis, for instance don’t turn off the lights in the living room if a person is present, more accurate than simple motion detection.
It works really well, I set up openHAB in a docker just for this. If I send messages to all of them at once, they’re not perfectly synchronized but it’s good enough for me.
Do you have a write up on how you did this, as I am looking to do the same thing, but using Surveillance Station on Synology. Also, what hardware are you using for yolo/tensorflow? Something with a hefty GPU?
You have instruction how you using yolo?
This is the current set up I use https://youtu.be/fwoonl5JKgo.
I was wondering if it’s possible to combine yolo?
I’m interested in your workflow with YOLO, I’ve got a computer setup to start messing with it and have YOLOv4 built. I’ve been running test images from my cameras through it but haven’t automated the results or detection yet. I have Blue Iris drop alert images onto the YOLO server, but I’m not sure how to have YOLO automatically scan them, or the best way to trigger it.
) These kind of problems where quite easy to fix by some new example images/trainings. After that it’s pretty good I have to say!