Implementation of visual people counting algorithms in embedded systems

Optimising the efficiency of HVAC systems represents a significant opportunity to reduce energy consumption in buildings and mitigate greenhouse gas emissions.This research evaluates low-resolution computer vision algorithms for occupancy detection on resource-constrained embedded systems.Our evaluation focuses specifically on the feasibility of deploying advanced AI object detection models on low-cost hardware platforms (under = C10) with varying computational capabilities.We systematically compared 45 different pre-trained object detection models using the COCO dataset.Among the models evaluated, those with the YOLO backbone proved to be the most suitable for this task.Quantitative analysis showed that YOLOv5n achieved a favourable balance between accuracy (AP50 = 0.944; AP50-95 = 0.584), model size (2.6 MB in RKNN format) and inference time.Performance tests on three embedded platforms -ESP32-CAM (microcontroller), Raspberry Pi Zero 2 W and Luckfox Pico Mini A (single-board computer) -revealed significant differences in inference speed, with hardware-accelerated solutions up to 10,000 times faster than software-only implementations.We have verified real-world applicability using our own ceiling-mounted wide-angle camera dataset.Future work will focus on developing a full hardware prototype, optimising the training dataset with AI-generated synthetic data, and implementing sensor fusion with audio signals for a multimodal approach.