Experiencing prosthetic vision with event-based sensors
| Autor: | Rückauer, B.J., Gerven, M.A.J. van, Potok, T.E., Schuman, C. |
|---|---|
| Přispěvatelé: | Potok, T.E., Schuman, C. |
| Rok vydání: | 2022 |
| Předmět: | |
| Zdroj: | Potok, T.E.; Schuman, C. (ed.), ICONS '22: Proceedings of the International Conference on Neuromorphic Systems, Article-20. New York, NY : Association for Computing Machinery (ACM) STARTPAGE=Article;ENDPAGE=20;TITLE=Potok, T.E.; Schuman, C. (ed.), ICONS '22: Proceedings of the International Conference on Neuromorphic Systems Potok, T.E.; Schuman, C. (ed.), ICONS '22: Proceedings of the International Conference on Neuromorphic Systems, pp. Article-20 |
| Popis: | Item does not contain fulltext Advances in materials science enable steadily increasing electrode counts for visual implants. On the algorithmic side, the fundamental problem remains how to encode a dense video feed using a low-dimensional set of stereotypical stimulation patterns. Bio-inspired retinal models have been proposed to improve the perceived light dots (phosphenes) resulting from electric stimulation. Here we present a system to evaluate prosthetic vision driven by a Dynamic Vision Sensor (DVS); a silicon retina modelling ON / OFF ganglion cells. This frame-free sensor signals pixel-wise changes in light intensity and is characterized by a high dynamic range and microsecond temporal resolution. We demonstrate how these retina-derived characteristics are beneficial in the presence of difficult lighting, motion blur, and for removal of cluttered redundant background. The system is tested in scenarios pertinent to prosthetic vision, such as recognition of letters, human gestures, pedestrians and cyclists, and compared against a setup with a conventional camera using edge-based stimulus generation. An AI agent trained to perform gesture recognition achieves 90% accuracy via simulated prosthetic vision driven by the DVS. ICONS '22: Proceedings of the International Conference on Neuromorphic Systems (Knoxville TN USA July 27-29, 2022) |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|