Maintaining accurate records for both the quantities and locations of inventory is vital when running any business operations efficiently and at scale. By leveraging new technologies such as AI and computer vision, items in warehouses, store shelves, and even a customer’s hand can be better managed and used to forecast changes demand. As demonstrated by the Zalmotek team, a tiny Arduino Nicla Vision board can be tasked with recognizing different types of containers and sending the resulting data to the cloud automatically.

The hardware itself was quite simple, as the Nicla Vision already contained the processor, camera, and connectivity required for the proof-of-concept. Once configured, Zalmotek used the OpenMV IDE to collect a large dataset featuring images of each type of item. Bounding boxes were then drawn using the Edge Impulse Studio, after which a FOMO-specific MobileNetV2 0.35 model was trained and could accurately determine the locations and quantities of objects in each test image.

Deploying the model was simple thanks to the OpenMV firmware export option, as it could be easily incorporated into the main Python script. In essence, the program continually gathers new images, passes them to the model, and gets the number of detected objects. Afterwards, these counts are published via the MQTT protocol to a cloud service for remote viewing.

You can read more about the proof of concept in much more detail here on the Edge Impulse blog.