Benchmarking End-to-End Deep and Feature Engineered Learning on an Embedded Smart Access Control System

Artificial intelligence has been extensively applied to address a range of security concerns, especially in access control systems. The paradigm has shifted from analyzing features meticulously extracted from images to deep learning of global patterns present in them. This paper presents end-to-end systems based on both prevailing technologies. Facial recognition models were deployed on a Raspberry Pi 4 board, interfaced via the MQTT protocol with an ESP32 module, which was tasked with electromagnetic locking and lighting controls. Access to the demonstrated area was granted only to authorized faces previously stored in the system. Additionally, the system featured liveness detection, preventing the malicious use of photographed images to bypass access. The benchmarks showed that CNN marginally outperformed the HOG-based algorithm on the testing dataset, although at higher computational costs. It was thus chosen for deployment in an actual environment and left unattended for one week. Consequently, the authorization accuracy reached 98.06% under specified conditions.