Arduino® Nicla Vision - Smart Trash Collector

Focoos AI

Arduino IDE (any version)

Nicla Vision

Z-Ant

Arduino CLI

<html><body><p>Keeping recyclables out of landfills often comes down to quick, reliable identification. <a href="https://github.com/ZantFoundation/Z-Ant/tree/main/examples/Nicla-SmartTrashCollector">Nicla-SmartTrashCollector </a>turns an Arduino Nicla Vision into a tiny edge model that classifies waste (plastic, paper, glass, organic, general) and lights the onboard RGB LED to guide sorting in real time. </p><p>Training and quantization of an object-detection model was done using the <a href="https://focoos.ai/">Focoos A </a>I platform, that allows exporting the quantized model in ONNX format. This format is digestible by the <a href="https://github.com/ZantFoundation/Z-Ant">Z-Ant </a>command line interface, that allows deployment of the quantized ONNX model on the Nicla Vision. </p><h3>How it works</h3><ol><li>The Nicla Vision captures 160×120 RGB frames from its onboard camera. </li><li>Frames are resized to 96×96 (NCHW) and fed to a 5-class classifier trained on Focoos AI. </li><li>Z-Ant handles ONNX conversion and a memory-mapped XIP flow for model weights. </li><li>The top class drives the RGB LED (yellow/blue/green/brown/grey) for instant feedback. </li></ol><h3>Dataset</h3><p>The dataset is publicly available at this link: https://www.kaggle.com/datasets/ceciliacasarella/waste-classification-dataset and it is a curated collection of images sampled from multiple publicly available sources to ensure balanced coverage across material categories. </p><p>Once the dataset `.zip` file is downloaded, it should be uploaded to the Focoos Platform using the “Add New Dataset” option. Select the Classification task type and, most importantly, choose the Classification Folder Layout to ensure the data is organized correctly for training. </p><p> </p><figure><img src="https://projects.arduinocontent.cc/b2daba24-8a6a-4466-aeb1-e6ef9736a84c.png"/><figcaption>Dataset Configuration</figcaption></figure><figure><img src="https://projects.arduinocontent.cc/030cda99-c75d-48cc-8654-46be10a84733.png"/><figcaption>Dataset Overview</figcaption></figure><p> </p><h3>Train → Export → Deploy</h3><ol><li>Focoos AI: train, quantize, and export model in ONNX version. </li><li>Z-Ant: `zig build lib …` to produce `libzant.a` from the ONNX model. </li><li>Arduino: include the library, compile the `.ino`, and flash internal firmware + XIP weights via DFU. </li></ol><h3>Details</h3><p>To understand how to train, quantize, and export a custom model using Focoos AI, these two links can be navigated: </p><ol><li><a href="https://focoosai.github.io/focoos/concepts/">https://focoosai.github.io/focoos/concepts/ </a></li><li><a href="https://github.com/MirkoCalvi/tiny_hack/blob/main/docs/quantize_classification.ipynb">https://github.com/MirkoCalvi/tiny_hack/blob/main/docs/quantize_classification.ipynb </a></li></ol><p>Training is carried out on the <strong>Focoos AI Platform with GPU support </strong>. </p><p>Once the dataset is uploaded, the next step is to define a new model. It is crucial to ensure that the task type remains set to “Classification” during this configuration. Select the nano model ( <strong>fai-cls-n-coco </strong>), which provides an optimal model size compatible with the Nicla Vision board’s RAM specifications. </p><figure><img src="https://projects.arduinocontent.cc/bdbb1934-b0bf-4143-b89d-623c52b3bb6f.png"/><figcaption>New model Settings</figcaption></figure><p>After configuring the dataset and model, a new training session can be initialized. The specific parameters used for this project are as follows: </p><h3>General Settings</h3><ol><li>Batch size: 8 </li></ol><ol><li>Evaluation period: 100 </li></ol><ol><li>Maximum iterations: 5000 </li></ol><ol><li><strong>Resolution: 96 </strong></li></ol><h3>Optimization</h3><ol><li>Optimizer: ADAMW </li></ol><ol><li>Learning rate: 0.00005 </li></ol><ol><li>Weight decay: 0.0001 </li></ol><ol><li>Scheduler: MULTISTEP </li></ol><ol><li>Early stopping: Enabled </li></ol><ol><li>Patience: 10 </li></ol><h3>Data Augmentation</h3><ol><li>Crop: Disabled </li><li>Horizontal Flip: 0.5 </li><li>Vertical Flip: 0.5 </li></ol><figure><img src="https://projects.arduinocontent.cc/846a0da5-b9c8-4bf9-a1f1-0cbe48b577e5.png"/><figcaption>Training Parameters</figcaption></figure><p> </p><p> </p><figure><img src="https://projects.arduinocontent.cc/d8998cba-93fe-41d0-b838-d3edb67ed2a0.png"/><figcaption>Trained model</figcaption></figure><p> </p><p>Once obtained the ONNX model, the first step is to clone the Z-Ant official repository. </p><p>The following are requirements to proceed further: </p><ol><li>Arduino core: `arduino:mbed_nicla` (tested 4.4.1) </li><li>`arduino-cli`, `dfu-util` (only if using XIP flashing script) </li><li>Zig 0.14.x </li></ol><p>Your ONNX model, for example, `my_model.onnx`, should be placed at `Z-Ant/datasets/models/my_model/my_model.onnx`. Then, you can issue the following commands to obtain a `libzant.a` file that will be placed on an Arduino library to be accessed by an Arduino .ino file. </p><ol><li><code class="code">zig build lib-gen  -Dmodel="YOUR_MODEL" -Dxip=true -Ddynamic -Ddo_export </code></li><li><code class="code">zig build lib-test -Dmodel="YOUR_MODEL" -Dxip=true -Ddynamic -Ddo_export </code></li><li><code class="code">zig build lib      -Dmodel="YOUR_MODEL" -Dtarget=thumb-freestanding -Dcpu=cortex_m7 -Dxip=true -Doptimize=ReleaseSmall </code></li></ol><p>The generated library should be copied to the Arduino Sketch using the following command: </p><code class="code">cp zig-out/YOUR_MODEL/libzant.a examples/Nicla-SmartTrashCollector/ZantLib/src/cortex-m7/</code><p>To compile and flash the Nicla, there are two bash scripts that should be slightly modified accordingly to your system. One is `flash_nicla_xip.sh` and the other is `compile_and_flash.sh`. Both of them are located in the project repository related to this post <a href="https://github.com/ZantFoundation/Z-Ant/tree/main/examples/Nicla-SmartTrashCollector">https://github.com/ZantFoundation/Z-Ant/tree/main/examples/Nicla-SmartTrashCollector </a></p><p>The same pipeline can adapt to other facilities or materials. The result is a compact, power-efficient vision node that’s easy to replicate and integrate. </p><p> </p></body></html>

Arduino® Nicla Vision - Smart Trash Collector

Turn an Arduino Nicla Vision into a tiny edge model that classifies waste (plastic, paper, glass, organic, general) and lights the onboard RGB LED to guide sorting in real time.