Nicla Vision custom image classification

What you’re building

Arduino’s Nicla Vision image-classification tutorial trains a custom TensorFlow Lite model in Edge Impulse and embeds it into OpenMV firmware for local camera inference. The first-party walkthrough demonstrates supervised classification of user-defined objects such as fruits; a person-detection project would require collecting and validating appropriate person and non-person classes rather than relying on a bundled detector. Hardware is the Nicla Vision, a micro-USB cable, an Edge Impulse account, and representative subjects or objects. The expected outcome is an OpenMV script that captures frames, runs the baked-in model, and reports class confidence without a cloud connection. The model must fit internal flash because the board lacks dedicated SRAM for runtime loading. Arduino recommends MobileNetV2 96x96 0.1, roughly 200 KB, and warns that larger-ROM model choices will likely not fit.

Key steps

1. Set up Nicla Vision in OpenMV IDE and create labeled image classes, including a diverse unknown or negative class.
2. Capture varied images across subjects, angles, lighting, and backgrounds rather than repeating one controlled specimen.
3. Upload the OpenMV dataset to Edge Impulse and retain about 20% as unseen test data.
4. Create a 48x48 RGB image impulse with transfer learning, generate features, and train MobileNetV2 96x96 0.1.
5. Evaluate the held-out test set for overfitting before exporting OpenMV Firmware.
6. Flash the generated Nicla Vision firmware, run ei_image_classification.py, and verify confidence on genuinely new scenes.