Create a Docker image on Raspberry Pi that runs an object detection model using LiteRT. The hardware setup consists of a Webcam connected to the Raspberry Pi board using a USB Hub. Fundamental ideas on AI model output tensor, LiteRT runtime, MQTT and Docker are discussed in Doulos course on Edge AI for Embedded Developers

Step 1: Clone the repository on either Raspberry Pi 3, 4 or 5

$git clone https://github.com/Doulos/Edge-AI-Sample

Cloning will create a new folder called Edge-AI-Sample

Change directory to this newly created folder.

Step 2:

Install Docker on Raspberry Pi using this command

$ curl -sSL https://get.docker.com | sh

Add your user to the docker group

$ sudo usermod -aG docker $USER

Step 3:

Create Docker image using build.

• The image is built locally on Raspberry Pi -3B and takes about 3-4 minutes.
• Change to the Edge-AI-Sample directory and issue the build command.

$sudo docker build -t yolov10-mqtt .

Step 4:

Check if the image is successfully created

$docker images

REPOSITORY       TAG       IMAGE ID       CREATED          SIZE
yolov10-mqtt     latest    c73cb468959d   7 minutes ago    405MB

Step 5: Run the Docker container created in step 3 in the background. Camera is connected on the /dev/video0

$sudo docker run --privileged -v /dev/video0:/dev/video0 yolov10-mqtt &

Step 6: Observe the output from the application.

• The output is a MQTT payload consisting of object label, confidence score and bounding box coordinates.
• Sample outputs with camera pointing at a person.

[3] 37984
pi@raspberrypi:~/Edge-AI-Sample$ INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Sending JSON: {"timestamp": 1766459874.8768485, "objects": {"class_id": 0, "label": "person", "confidence": 0.793, "bbox": [0.09138107299804688, 0.07930111885070801, 0.6551454067230225, 0.9981493949890137]}}
Inference published to MQTT topic
Logged to DB: person
MQTT client disconnected and loop stopped.
================================================================================
ID   | Timestamp            | Label      | Conf   | Coordinates (x1, y1, x2, y2)
--------------------------------------------------------------------------------
1    | 2025-12-23 03:17:54  | person     | 0.79   | (0.09138107299804688, 0.07930111885070801,
     |                      |            |        |  0.6551454067230225, 0.9981493949890137)
================================================================================

Step 7:

• Change the MQTT topic name in the application code (object-detection-yolov10.py) and subscribe to the topic on another computer to view the output from the object detection model.
• Try updating the code to give continuous inference. Right now, it is setup to provide only one inference at start.
• Use the SQLite database content to create a dashboard using Streamlit