Speaker
Description
Reverse vending machines encourage recycling, but require automatic container classification. This is typically achieved by reading barcodes or using computer vision, which can fail due to unreadable codes or poor lighting. Recently, an active acoustic approach using sound waves has been proposed as an alternative, offering low computational cost and minimal maintenance. However, its performance depends on the acoustic environment. This study investigates how enclosure geometry and wall absorption influence classification accuracy. Acoustic impulse responses were measured using exponential sine sweeps, with an omnidirectional parametric loudspeaker generating ultrasonic and audible waves via the parametric acoustic array effect. The data trained machine learning and deep learning models to classify plastic, metal cans, glass, and cartboard-tetrabrick. The tests were performed in a mini-reverberation chamber and in a shoebox enclosure with and without wall-mounted absorbers. The highest accuracy was achieved in the mini-reverberation chamber (96%), highlighting the benefits of diffuse and reverberant environments.