Low-Power human face tracking system for art installations at low ambient light
Michael N. Mayer, Guan‐Ru Chen, Tzu-Feng Chiu, Shu-Ciang Chen · 2023
This paper aims to create an interactive human-machine art exhibition solution for a light and shadow art installation through embedded edge computing. Originally, this device used a microcontroller for simple microcontrol of the stepper motor, which did not meet the artists’ needs. Therefore, we added a high-sensitivity lens module and an embedded NVIDIA edge computing platform to the device and equipped it with an interactive method for efficient real-time face recognition and tracking at night through YOLOv4 tiny. We improve the YOLOv4-tiny-3l model by using various hyperparameter adjustments and image preprocessing methods to increase the diversity and generalization capability of the dataset through self-training. This allows the YOLOv4-tiny-3l network structure model to improve face recognition speed by 5–10% and average accuracy by 1–5% at a training resolution of 320×320 compared to the YOLOv4-tiny network structure at a training resolution of 416×416.