<![CDATA[Manual attendance recording remains inefficient and may enable proxy attendance, while cloud-based facial recognition introduces network dependency, latency, and biometric-data privacy concerns. Edge processing offers a local alternative, but limited computing resources require pipeline optimization. This study conducted a descriptive single-run evaluation of five optimization strategies on a Raspberry Pi 5: frame downscaling, static region of interest (ROI), prototype caching, sample interval tuning, and face detector comparison. The pipeline used FaceNet embeddings, cosine-similarity-based open-set recognition, and track-based temporal confirmation. Evaluation involved three prerecorded 1920 × 1080 classroom videos totaling 16 min 23 s, an enrollment database of 18 students, and 53 student-video attendance instances. Frame downscaling at 0.75 reduced runtime from 1,274.83 s to 947.72 s, equivalent to 25.66%, while confirmed attendance instances decreased slightly from 47 to 46. Effective sampled-frame throughput increased by 34.62%, and average RAM usage decreased by 6.86%, although CPU utilization and peak temperature increased. Prototype caching reduced preparation time but had limited effect on end-to-end runtime, whereas static ROI, sampling intervals above 5 s, and YuNet did not improve overall pipeline performance. Frame downscaling was therefore the most promising strategy in the evaluated setup. However, repeated trials and independent classroom-video testing are required before generalizing the results or considering the pipeline deployment-ready.]]>
Copyrights © 2026
