Smart Counting: PIR-Gated YOLOv۸ and ESP۳۲-CAM for Energy-Efficient Edge AI

Accurate and energy-efficient people counting is essential for smart buildings, retail analytics, and occupancy management, yet existing solutions force a trade-off: continuous-stream vision systems achieve high accuracy at the cost of bandwidth and energy, while low-power sensor-based alternatives sacrifice directional precision. This paper presents a three-tier Edge AI architecture that bridges this gap. An ESP۳۲-CAM edge node, gated by a passive infrared sensor, remains in a low-power idle state and transmits JPEG snapshots only upon motion detection, thereby eliminating redundant data transmission. Snapshots are processed on a local edge server running YOLOv۸n for person detection coupled with a SORT-style centroid tracker and line-crossing logic that enables bidirectional entry and exit counting. A time-series occupancy log with an integrated query interface supports retrospective analytics. Experimental evaluation over thirty controlled entry and exit events demonstrates an overall counting accuracy of ۹۶.۶۷ percent, with ۱۰۰ percent entry accuracy and ۹۲.۸۶ percent exit accuracy. The PIR-gated architecture achieves an effective power consumption of ۴۴۶.۳۱ milliwatts and requires only ۱۵۳.۸۴ kilobits per second of bandwidth, representing a ۷۶.۳۳ percent reduction compared to equivalent MJPEG streaming. Estimated daily energy consumption of ۱۰.۷۱ watt-hours enables sustained operation on limited power sources. These findings confirm that the proposed design delivers a practical, privacy-preserving, and cost-effective solution for indoor bidirectional people counting.

کلیدواژه ها:

Edge AI ، People counting ، ESP۳۲-CAM ، YOLOv۸ ، PIR sensor