For recycling and non-recycling classes, refer to /dataset. The directory includes train and test datasets for recycling and non-recycling, as well as a folder of empty images. These images can be automatically split into train/test sets on Edge Impulse.
To use already augmented training and test data, refer to /dataset_augmented. To generate your own augmented data, run
./cache4trash/data_augmentation.py and define the current_path as the directory where you want to generate the augmented data.
Upload both train and test data for all three classes (recycling, non-recycling, and empty) to Edge Impulse as an image classification task. Create an Impulse for image classification (or image transfer learning) and run the model. For better results, configure the EON Tuner to the Arduino Nicla Vision device and run it. Choose a model with high accuracy that also fits on the Nicla Vision (< 220K).
To deploy our already-trained model to Arduino Nicla Vision, first clone this repository, then simply connect the Nicla to the computer and open OpenMV IDE. Open the
./cache-4-trash/cache4trash/final_nicla_cache4trash.py
script from the cache-4-trash repository.
Within OpenMV IDE, go to -> Tools -> Run Bootloader (Load Firmware) to load our firmware into the Nicla. select the pathway on your local computer that leads to
./cache-4-trash/open_mv_firmware/edge_impulse_firmware_arduino_nicla_vision.bin.
Select "Erase internal file system" and click "Run".
After the light flashes and the upload is completed, press the green start button, and the feedback loop should start working. Use the Nicla to capture images of common objects and in the Serial terminal, the probability of whether that item is recyclable or not recyclable will be printed.
For the full hardware setup, please see the circuit diagram below.
Connect the corresponding Nicla pins to the corresponding LED lights. Additionally, add manual overwrite buttons to PF_3 and PG_12 pins. Then, Attach photo-transistors (next to the two LED lights), H-bridge (with DC motor and 9V battery connected), and an external power source (or computer power cord) to the corresponding places on the Arduino MKR Zero. Additionally, compile and deploy the Arduino code in
./cache-4-trash/cache4trash/cache4trash.ino to the Arduino MKR Zero.
After all, is completed, the full mechanism for cache 4 trash is built, and attaching it to a cache system on top of a trash can while attaching the motor to an electric seesaw similar to the diagram below will result in the full hardware deployment of the cache 4 trash system.
Following model deployment through firmware installation and OpenMV IDE, click the green button while Nicla is connected to your machine and you are currently on the ./cache-4-trash/cache4trash/final_nicla_cache4trash.py script. This should start the inference feedback loop. Please open the Serial monitor to see the inference results of past snapshots.