AI-Powered Smart Doorbell: An ESP32-CAM and Cloud-Based Prototype for Visitor Recognition and Staged Notifications

Low-cost camera-based doorbells can provide visitor information through embedded image capture and remote inference, but recognition quality and user-visible delay must be evaluated together. This paper presents an ESP32-CAM prototype that captures an image when a visitor presses a button and transfers it to a cloud backend. RetinaFace detects and aligns faces, and ArcFace produces 512-dimensional embeddings through the DeepFace framework. Stored profiles in Firebase Firestore support known/unknown classification, while a mobile application presents an initial image notification followed by a recognition update. Evaluation uses 50 classification events and 14 timing pairs. The results give 91.7% precision, 100% recall, 95.7% F1-score, and 94.0% accuracy. Three of 17 unknown events were incorrectly classified as known, corresponding to a 17.6% unknown-event false-positive rate. Mean initial-notification and recognition-endpoint times were 4.70 and 8.05 s, respectively. The listed hardware components total CAD 29.98, excluding deployment and recurring service costs. These findings demonstrate feasibility of the integrated sensing and notification pipeline while identifying false acceptance, incomplete experimental documentation, and cloud-dependent delay as limitations requiring further validation before residential deployment.