Colonnier, F., Seeralan, A., & Zhu, L. (2023). Event-Based Visual Sensing for Human Motion Detection and Classification at Various Distances. Lecture Notes in Computer Science, 75–88. https://doi.org/10.1007/978-3-031-26431-3_7
Abstract:
In Human Research and Rescue scenarios, it is useful to be able to distinguish persons in distress from rescuers. Assuming people requiring help would wave to attract attention, human motion is thus a significant cue to identify person in needs.
Therefore, in this paper, we aim at detecting and classifying human motion at different depths with low resolution. The task is fulfilled thanks to an event-based sensor and a Spiking Neural Network (SNN). The event-based sensor has been chosen as a suitable device to register motion specifically. While SNN is appropriate to process the event-based data, it is also a suitable algorithm to be implemented in low-power neuromorphic device, allowing for a longer operating time. In this study, we gather new data with similar classes to the IBM DVS Gesture dataset at various distances. We show we can achieve an accuracy up to 90.7% on a validation set obtained at different depths and lighting conditions from the training set. We also show that having an Region of Interest detection leads to better accuracy compare to a full frame model.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the National Research Foundation - Research, Innovation and Enterprise 2020 (RIE2020)
Grant Reference no. : A1687b0033