Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: Focus: Deep Learning in Electromagnetics Research

AKPIK 4.3: Vortrag

Dienstag, 10. März 2026, 15:00–15:15, BEY/0127

Optical Human Action Recognition - Less can be more? — •Maximilian Zier, Stefan Sinzinger, Kathy Lüdge, and Lina Jaurigue — Technische Universität Ilmenau, Ilmenau, Germany

Automated recognition of human actions is becoming more relevant due to applications in areas such as surveillance and autonomous driving. Modern neural networks achieve nearly perfect classification accuracies across various action datasets. However, they rely on complex feature extraction methods that lead to long training times and significant computational demands. Focusing on sustainability and efficiency, Reservoir computing systems aim to deliver similar performance combined with reduced computational effort by only training the output layer. An optically implemented reservoir offers the prospect of processing at the speed of light and nearly unlimited scalability due to inherent parallelism. In this contribution, we present results of a human action recognition task using a hybrid opto-electronic set-up based on [1]. In contrast to previous works, we forgo common preprocessing and feature extraction methods and use raw video data as input to the reservoir. Our system lags behind large neural networks in terms of classification accuracy, but has very low hardware requirements. Additionally, we reduce the length of the video sequences to one second, thereby using less input data to perform classification than previous works.
[1] Antonik, P., Marsal, N., Brunner, D. et al., Nat Mach Intell 1, 530-537 (2019)

Keywords: Reservoir Computing; Optic; Human Action Recognition