Design and Development of SmartGuard: A YOLOv8-Based Home Intruder Detection and Early Warning System

Abstract

Residential surveillance remains constrained by passive monitoring, continuous human supervision, subscription-based services, and limited interoperability. This study developed and evaluated SmartGuard as a modular residential-surveillance architecture integrating a fine-tuned YOLOv8n person detector, dual-camera round-robin monitoring, three-frame temporal validation, Telegram-based warning delivery, event logging, and optional local face matching. SmartGuard detects persons; when the optional identity component does not match a registered occupant, the system issues a provisional unrecognized-person warning rather than independently determining that the individual is legally or contextually an intruder. The system was developed through the Planning, Design, Development, and Testing stages of an Agile process and implemented as a five-layer architecture. The detector was trained on a merged Roboflow Universe corpus and evaluated on 839 held-out test images. It achieved 94.2% precision, 93.0% recall, an F1-score of 0.94, mAP@0.5 of 96.6%, and mAP@0.5:0.95 of 74.6%. The reported 4.0 ms represents model-only inference on a Tesla T4 GPU and does not represent end-to-end latency on the intended Intel Core i5 deployment platform. The available records did not specify the exact number of annotated person instances, independently observed entry events, or repeated latency measurements. Black-box testing showed that all five functional modules met their predefined criteria. In a single-residence trial, 520 alert screenshots were generated; 504 were evaluable and 503 were correct, yielding 99.8% post-alert precision after excluding 16 indeterminate cases. This value does not estimate field recall, overall accuracy, missed-person events, or operational security effectiveness.