Comparison of Artificial and Spiking Neural Networks for Ambient-Assisted Living

In assisted living environments, various situations may arise where a person falls or is otherwise injured and is unable to call for help on their own. In such situations, it is necessary to quickly identify the problem and take appropriate action, such as calling for help. This can be supported or even automated by using vision-based AI systems. In this context, we investigated and evaluated different AI solutions for rapid human action recognition. More specifically, we trained and compared artificial neural networks (ANN) in combination with frame-based cameras to a processing pipeline using spiking neural networks (SNN) and event-based cameras. For the SNNs, we further distinguished and compared two models, which we simulated in software and implemented on event-based hardware. The SNNs feature various layer types, e.g. fully-connected, spiking convolutions and recurrent. The implementation on event-based hardware was compared to GPU-based embedded hardware for artificial neural networks. The comparison was made primarily with regard to the highest possible energy efficiency in order to enable battery-powered vision systems in the future that can be used flexibly not only in assisted living, but also in industrial and smart city applications. The networks were constructed in such a way that they achieve a similar classification accuracy, which was measured on our own dataset specifically recorded for the project.