Running a custom AI model on Arduino UNO Q: Steps for creating an image classification model using Edge Impulse

1-2. Preparations

• Arduino UNO Q board (pre-configured and ready to connect to a PC)
• Arduino App Lab (already installed on PC)
• USB camera (used for collecting training data and inference)
• Arduino account (required to log in to Edge Impulse)

2-2. Open the Image Classification Brick

Once connected, select Bricks from the left menu, and then select Image Classification from the list.

Next, select the AI models tab and then select the Train new AI model button at the bottom.

Classification Brick Overview

AI models tab and Train new AI model button

You will be prompted to register and log in to your Qualcomm account. Once you log in successfully, your Edge Impulse user page will open in your browser. Model creation will take place on this website.

2-3. Create a project

Select "Create new project" in the upper right corner, enter a project name, and create the project.

Edge Impulse user page

Project creation dialog

2-4. Configure the target device

Select "Target" in the upper right corner, then select "Arduino UNO Q (Qualcomm QRB2210)" from the list of target devices and save.

Target device selection 1

Target device selection 2

Also, under "Project info" in the bottom right, select "One label per data item" from the "Labeling method" (in this case, since it's image classification, we'll assign one label to each image).

Labeling method settings

3-1. Connect the data acquisition device.

Select "Collect new data" from "Getting started".

Select a device for data entry. You can also scan the QR code on the left with your smartphone and take a photo with your smartphone's camera to enter data. This time, we will select "Connect to your computer" in the center and enter data using a USB camera connected to a PC.

Select device 1 for collecting new data.

Device selection 2 for collecting new data

Once the connection is complete, select "Collecting images?" and then select "Give access to the camera" to grant permission to the connected camera.

Select device 3 for collecting new data.

Select device 4 for collecting new data.

3-2. Collect images of "mouse"

First, we'll collect an image of the mouse. Select the Label button at the top, label it 'mouse', and then press Capture to record the image of the mouse.

Capture approximately 30 images of the mouse cursor, changing the angle and background each time.

Mouse capture screen

Label input

3-3. Collect images of "scissors"

Similarly, label the image 'scissors' and capture about 30 images of scissors, varying the angle and background.

Labeling and capturing scissors 1

Labeling and capturing scissors 2

3-4. Collect images of "others"

Next, capture about 30 images other than scissors and mice and label them as 'others'. By training the model with items other than those to be classified as "OTHERS," you can reduce false positives.

Labeling and capturing others 1

Labeling and capturing others 2

4-1. Create an Impulse

From the Impulse design menu on the left, select Create impulse.

• Add an image to the Add processing block.

Create impulse: Add screen and blocks 1

Create Impulse: Adding screen and blocks 2

• Add the Transfer Learning (Images) block to the Add learning block.​
• Select Save Impulse to save.

Create impulse: Add screen and blocks 3

Adding screen and blocks to Create Impulse 4

Impulse after configuration is complete

4-2. Generate features

Select "Image" from the Impulse design menu on the left, and then select the "Generate features" tab from the menu at the top. Select the "Generate features" button on the screen that appears.

Executing GenerateFeatures 1

Confirm that the job is completed in the log below.

Run 2 of Generate features

4-3. Training the Model (Transfer Learning)

From the Impulse design menu on the left, select Transfer learning, and then select Save & train at the bottom.

Learning settings and Save & Train

The job is complete when "Job completed" is displayed in the log in the upper right corner. Model performance is also measured, and results such as Accuracy and confusion matrix are displayed at the bottom of the log.

Learning results (Accuracy and confusion matrix)

6-2. Create an app

Select "My Apps" from the top left menu, then select "Create new app+". Enter a project name and select "Create new".

Create a new app

6-3. Add Brick

Select "Add bricks" from the top left menu. Select "Video Image Classification" from the list on the left, and then select the project you just created from the list of AI models on the right and add it using "Add brick".

Adding a Video Image Classification Brick

Adding a Video Image Classification Brick

6-4. Rewrite main.py

Replace the contents of 'main.py' located under 'python' in the left-hand menu with the following:

from arduino.app_utils import App
from arduino.app_bricks.video_imageclassification import VideoImageClassification

classification_stream = VideoImageClassification()

def detected(label):
    print(f"Detected: {label}")

classification_stream.on_detect("mouse", lambda: detected("mouse"))
classification_stream.on_detect("scissors", lambda: detected("scissors"))
classification_stream.on_detect("others", lambda: detected("others"))

App.run()

6-5. Run the app.

Start the demo by clicking the Run button in the upper right corner.

After startup, switching the log display to Python will classify objects captured by the USB camera. In this case, since we've trained it to recognize the PC mouse, pointing the camera at the mouse will trigger a log message indicating mouse detection.

Mouse detection log