Lightweight deep learning for autonomous human counting system on low-cost hardware

Abstract

Accurate and efficient human counting is essen- tial for optimizing public transportation and advancing smart city infrastructure. This paper evaluates pro- posed lightweight deep learning models for autonomous human counting system on low-cost hardware, ensur- ing real-time monitoring and enhanced operational ef- ficiency. While existing methods, such as DeepSORT, Kalman Filters, and YOLO variants, are often im- plemented on high-end hardware, they typically prior- itize accuracy over computational efficiency. Few ob- ject detection and tracking techniques can run in real- time on low-end hardware. This work advances the field by utilizing optimized deep learning models suit- able for embedded systems with constrained resources. Specifically, fine-tuned YOLOv8 is employed for head detection, combined with ByteTrack for robust track- ing, outperforming YOLOv5 and YOLOv11 in accu- racy and efficiency. Archiving the 15 FPS and more then 90% accuracy on the real environment deployment on both RISC-V architecture with an integrated NPU (MaixCAM) and ARM v8 (Raspberry Pi), The pro- posed system demonstrates its suitability for real-time, cost-effective, and scalable autonomous human count- ing in public transit environments.