All About Automations

[SmartPad]

Amen!

 

[SmartPad]

We will pop back here and update everyone with a link as soon as we hear that the Items are manufactured and ready to ship, hopefully it will be in around 3-4 weeks

Hello everyone,

As promised, just to let everyone know that our range of new South African standard Zigbee, ESPHome and Tasmota devices have landed and are available in our store.

 

[valiente]

Hello everyone,

As promised, just to let everyone know that our range of new South African standard Zigbee, ESPHome and Tasmota devices have landed and are available in our store.

Plugs have been amazing btw!

 

[valiente]

Wanted to share my latest automation/playground, we have a shared vehicle gate for 3 units, sometimes this gate is left open (Press button while gate is closing).

Now, I already have a little Sonoff Gate opener which we all use to open this gate remotely, but I needed to tell the state of the gate, and in case it is left open, we can be alerted.

I played with the idea of adding a Reed switch (Wireless or Wired) or replacing the Sonoff with a Shelly that has an Input and then read the state from the Gate motor directly - Each option had problems and would have been more inconvenience caused that comfort gained...

Then I read somewhere that Frigate 0.17 will come with specific image classification, allowing recognizing WHO is that person or dog.

This inspired me to create my Classification Model, trained on almost 1000 photos of my gate, this model runs inside a Docker container with an API that accepts an Image and spits out Open or Closed

Started with the full resolution 2k image, which worked great during the day, but at night caused some false positives.

I then went onto crop the source image at a specific location, where the gate is fully closed - Cropped all my training images, adjusted my model to this smaller 64x64 image size and retrained:
Closed: Open:

Inference Speed is sitting at 4ms - 6ms to get the following reply:

{
   "status": "closed",
   "confidence": 1,
   "probabilities": {
      "closed": 1,
      "open": 2.2574055378103664e-16
   },
   "latency_ms": 5.72,
   "providers": [
      "OpenVINOExecutionProvider",
      "CPUExecutionProvider"
   ],
   "model": "gate_mobilenetv3.onnx"
}

The entire process:

1. Capturing a Screenshot from the Camera
2. Saving to Disk
3. Sending that to the Container
4. Saving the result to HA Sensor

All of that takes less than half a second:



Now I have a pretty accurate way of telling if the Gate is OPEN or CLOSED with no longer than 2 seconds of delay - Since I am running a classification every 2 seconds.

I am also still taking more photos for my training, as with different lighting, rain or night, the outcome might be different - I want to get to 10 000 training images.

 

[RoganDawes]

Wanted to share my latest automation/playground, we have a shared vehicle gate for 3 units, sometimes this gate is left open (Press button while gate is closing).

Now, I already have a little Sonoff Gate opener which we all use to open this gate remotely, but I needed to tell the state of the gate, and in case it is left open, we can be alerted.

I played with the idea of adding a Reed switch (Wireless or Wired) or replacing the Sonoff with a Shelly that has an Input and then read the state from the Gate motor directly - Each option had problems and would have been more inconvenience caused that comfort gained...

Then I read somewhere that Frigate 0.17 will come with specific image classification, allowing recognizing WHO is that person or dog.

This inspired me to create my Classification Model, trained on almost 1000 photos of my gate, this model runs inside a Docker container with an API that accepts an Image and spits out Open or Closed

Started with the full resolution 2k image, which worked great during the day, but at night caused some false positives.

I then went onto crop the source image at a specific location, where the gate is fully closed - Cropped all my training images, adjusted my model to this smaller 64x64 image size and retrained:
Closed: View attachment 1852094 Open: View attachment 1852095

Inference Speed is sitting at 4ms - 6ms to get the following reply:

{
   "status": "closed",
   "confidence": 1,
   "probabilities": {
      "closed": 1,
      "open": 2.2574055378103664e-16
   },
   "latency_ms": 5.72,
   "providers": [
      "OpenVINOExecutionProvider",
      "CPUExecutionProvider"
   ],
   "model": "gate_mobilenetv3.onnx"
}

The entire process:

1. Capturing a Screenshot from the Camera
2. Saving to Disk
3. Sending that to the Container
4. Saving the result to HA Sensor

All of that takes less than half a second:
View attachment 1852097


Now I have a pretty accurate way of telling if the Gate is OPEN or CLOSED with no longer than 2 seconds of delay - Since I am running a classification every 2 seconds.

I am also still taking more photos for my training, as with different lighting, rain or night, the outcome might be different - I want to get to 10 000 training images.

My gate controller monitors the status LED output of the Centurion D5 (and probably others), and determines the state of the gate based on the flashing of the LED. Off - gate is closed. On - gate is stationary, open. Flashing fast - gate is opening. Flashing slowly - gate is closing. Also triggers the gate using an opto-transistor to open/close/stop, etc.

 

[SmartPad]

Plugs have been amazing btw!

Hey valiente,

Thanks so much for the positive feedback