Using KubeAI Application Nucleus for edge

Introduction

I have a real world scenario requirement to implement an environmental sensor to count vehicles driving infront of my balcony. To do this my inital idea was to set up a camera and attach it to a Coral/TPU powered Rasperry Pi to catch the stream from the camera and count vehicles.

As we all know, RPis and also TPUs arent available and anything then a good deal at the moment. Good timing KAN jumped in and offers ready to use AI Skills on Kubernetes running on edge.

What is covered or what is the TL;DR here ?

We use a Kubernetes Cluster to deploy KAN into it and use the ready to use AI Skills to connect our Videostream from the camera to detect vehicles.

Parts involed:

• Reolink RC-510 Outdoor webcam with PoE
• Kubernetes Cluster deployed in Azure

What do we want to do ?

Stream our Camera feed via RTSP to a remote Kuberentes Cluster or Kubernetes system running on the edge for local processing. The long term goal is to capture the stream, seperate lanes and directions and count the amount of objects to expose this via API as environmental sensor.

The KAN Portal can be easily deployed via Helm chart and some lines of code into a running Kubernetes Cluster. After that you just need to add your source and processing system and thats it.

I deployed KAN via the following documentation

So this is our first output after attaching the RTSP stream to our KAN deployment.

For this you need to create a workflow (AI Skill) with the following parts:

• Skill name
• Device: CPU (your Kubernetes Cluster vCPU)
• Camera: the registered link to your RTSP camera
• and finally deploy the whole solution

You could attach a IoT Hub connection to catch the output of the detection/skill

As you can see the probability needs to be tuned as there are currently some issues with detection of static objects. But when you keep in mind that this doenst need a TPU or GPU support and is completely low-code, I’m very surprised about the outcome.

Even tho the results are pretty good, I dont want to handle a fully blown Kuberentes Cluster. If there is ongoing progress on the AI Skills built into KAN, we can test this again but for the moment I will probably switch to another platform.