An Edge-Deployable Privacy-Preserving Occupancy Detection System for Public Transport

Accurate occupancy estimation and passenger counting are critical for ensuring safety, efficiency, and effective management in public transportation systems. In this work, we present a privacy-preserving passenger counting system based on a Thermopile array image sensor (TPAS). By leveraging machine learning methods, the system extracts occupancy information from the spatial distribution of detected thermal radiation, without capturing personally identifiable information. We evaluated both the YOLOv8n detection model and a custom-designed TinyConvNet model optimized for deployment on edge computing platforms. On our self-collected BusLab dataset, YOLOv8n achieved a precision of 92.7% at an IoU threshold of 0.7, while TinyConvNet reached 78.8%. We further investigated the deployability of both models on edge devices and assessed their inference performance. The results demonstrate that both networks can be successfully deployed on hardware platforms installed in buses, enabling real-time monitoring while adhering to privacy-preserving principles.