Myers, Karl and Secco, Emanuele Lindo (2020) A Low-Cost Embedded Computer Vision System for the Classification of Recyclable Objects. In: Soft Computing Research Society and Congress on Intelligent Systems (CIS) 2020, September 05-06, 2020, Virtual Format. (Accepted for Publication)
Preview |
Text
Myers_Secco.pdf - Accepted Version Download (872kB) | Preview |
Abstract
Due to rapid urbanization, increasing population and industrialization, there has been a sharp rise in solid waste pollution across the globe. Here we present a novel solution to this inefficiency, by the use of embedded Computer Vision (CV) in the Material Recovery Facilities (MRF).
The proposed architecture employs software (i.e. Tensorflow and OpenCV) and hardware (i.e. Raspberry Pi) as an embedded platform in order to classify daily life objects according to their visual aspect. The CV system is trained using mod-ules contained within the TensorFlow API with two datasets, namely the Trash-Net and a combination of the TrashNet and a set of web images.
This solution allows greater accuracy, with a baseline performance of 90% which drops to 70% when deployed on the embedded platform, due to the quality of the images taken by an integrated camera for the real-time classification. The speed results are also promising with a baseline speed of 10 FPS at simulation level, which drops to 1. 4fps when running on the platform.
Such a system is cheap at less than £ 100, it is perfectly adequate to be used to identify recyclables in the MRF for sorting.
| Item Type: | Conference or Workshop Item (Paper) |
|---|---|
| Additional Information and Comments: | This is the authors' version of a paper that has been accepted for presentation at the Soft Computing Research Society and Congress on Intelligent Systems (CIS) 2020. |
| Faculty / Department: | Faculty of Human and Digital Sciences > Mathematics and Computer Science |
| Depositing User: | Emanuele Secco |
| Date Deposited: | 10 Aug 2020 10:16 |
| Last Modified: | 07 Sep 2020 00:15 |
| URI: | https://hira.hope.ac.uk/id/eprint/3118 |
Actions (login required)
 |
View Item |
Altmetric
Altmetric
