CPU-Optimized Real-Time Face Recognition For Automated Attendance Management: A Mediapipe-Based Approach With Transparent Audit Logging

Attendance monitoring constitutes a foundational administrative task in educational institutions, corporate environments, and regulated workplaces. Beyond simple headcount verification, attendance records serve compliance, performance analytics, resource planning, and legal accountability functions. Despite this significance, the predominant method of attendance collection in many institutions worldwide remains manual: paper registers, oral roll calls, or spreadsheet entries managed by instructors. These approaches are labor-intensive, error-prone, and susceptible to proxy attendance—a well-documented form of misconduct in which a student signs or responds on behalf of an absent peer [1].

Efforts to automate attendance collection have followed several technological trajectories. Radio Frequency Identification (RFID) card-based systems replace paper with electronic proximity scanning but retain the fundamental vulnerability of proxy attendance, since the token—not the person—is verified [2]. Fingerprint biometric systems address this shortcoming by verifying the individual directly, yet introduce hardware acquisition costs, hygiene concerns, and accessibility issues for individuals with dermatological conditions [3]. Cloud-connected artificial intelligence (AI) systems leveraging deep convolutional neural networks (CNNs) achieve state-of-the-art recognition accuracy but require continuous internet connectivity, expose sensitive biometric data to third-party servers, and typically demand GPU-accelerated inference infrastructure that is cost-prohibitive for resource-constrained institutions [4].

A meaningful gap therefore exists in the literature: the absence of an attendance system that is simultaneously biometrically secure, CPU-deployable without cloud dependency, transparent in its audit logging, and calibratable to institution-specific accuracy requirements. The present work addresses this gap by proposing a system built upon MediaPipe Face Mesh [5]—an open-source facial landmark detector from Google Research optimized for real-time CPU performance—combined with a lightweight embedding pipeline and threshold-calibrated cosine similarity matching.

The principal contributions of this work are as follows:

•  A complete, end-to-end face recognition attendance pipeline optimized exclusively for CPU inference, requiring no GPU or cloud connectivity, validated on commodity hardware.

•  A threshold calibration methodology for cosine similarity-based face verification using Receiver Operating Characteristic (ROC) analysis to balance FAR and FRR according to institutional requirements.

•  A transparent, human-readable CSV logging architecture with built-in duplicate suppression, enabling straightforward audit and compliance verification.

•  An empirical evaluation across N = 70 participants in three operational environments, with statistical validation of accuracy, latency, scalability, and illumination robustness.

The remainder of this paper is organized as follows: Section II surveys related work. Section III describes the proposed methodology. Section IV presents the experimental evaluation. Section V discusses findings and limitations. Section VI concludes.

RELATED WORK

The evolution of attendance management systems can be characterized across four technology generations: manual paper-based, token-based electronic, contact biometric, and vision-based AI systems. Table I summarizes a comparative analysis of representative systems across these generations, positioning the proposed approach within the solution space.

† RFID accuracy denotes token-read success rate; identity verification accuracy is lower due to card-sharing vulnerability.

A. Manual and Token-Based Systems

Paper-based roll calls and instructor-managed registers represent the performance baseline. Bowser and Calandra [1] documented error rates of 8–12% in manually maintained records over a semester-length study. RFID-based systems, reviewed by Musa et al. [2], reduce transcription error but transfer the authentication problem from identity to token possession. Proxy attendance via shared RFID cards remains trivially feasible.

B. Contact Biometric Systems

Fingerprint attendance systems directly authenticate the individual and have been widely deployed in academic institutions [3]. However, Jain et al. [6] identify three systemic limitations: (i) high-quality capture requires controlled conditions, leading to elevated FRR operationally; (ii) certain demographic groups exhibit low-contrast ridges causing systematic failures; and (iii) shared sensor surfaces raise hygiene concerns, a consideration amplified by the COVID-19 pandemic. Iris recognition systems address accuracy limitations at prohibitive hardware costs, typically USD 400–800 per terminal [7].

C. Vision-Based and Deep Learning Systems

Deep CNN models—FaceNet [8], ArcFace [9], DeepFace—have driven recognition accuracy to near-human levels on benchmark datasets. Schroff et al. [8] reported 99.63% accuracy on Labeled Faces in the Wild (LFW) using FaceNet triplet-loss embeddings. Kumar and Singh [10] demonstrated 98.2% accuracy with a cloud-connected FaceNet deployment across 120 students. However, these systems impose GPU inference servers, broadband connectivity, and third-party biometric storage—raising data sovereignty concerns under regulations such as GDPR [11].

D. Lightweight and Edge Deployments

Lugaresi et al. [5] introduced MediaPipe, a cross-platform framework providing real-time ML inference optimized for mobile and edge devices. MediaPipe Face Mesh achieves sub-50 ms landmark detection on modern CPU hardware. Zhang et al. [12] applied MediaPipe landmarks to lightweight classifiers for emotion recognition, achieving 94.1% accuracy on CPU-only hardware. Attendance-specific applications of MediaPipe remain limited in peer-reviewed literature, representing the primary research opportunity addressed herein.

PROPOSED METHODOLOGY

A. System Architecture Overview

The proposed system is structured as a four-layer pipeline (Fig. 1). Layer 1 (Input) handles real-time video frame acquisition. Layer 2 (Feature Extraction) applies MediaPipe Face Mesh to yield normalized 468-point landmark representation. Layer 3 (Verification) generates a 128-dimensional embedding and computes cosine similarity against enrolled references. Layer 4 (Persistence) manages CSV attendance logging with timestamp assignment, duplicate prevention, and session management.

Fig. 1 – Proposed Four-Layer System Architecture

B. MediaPipe Face Mesh Integration

MediaPipe Face Mesh [5] applies a two-stage inference pipeline per video frame. The first stage employs a lightweight face detector to localize candidate regions. The second stage applies a 468-landmark regression network to the localized crop, producing three-dimensional (x, y, z) coordinates for each landmark normalized to [0, 1]. For the proposed system, 2D (x, y) coordinates of all 468 landmarks are assembled into a 936-dimensional feature vector. To achieve translation and scale invariance, each coordinate is centered at the facial bounding box centroid and normalized by the inter-ocular distance d_io (Euclidean distance between outer eye corners, landmarks 33 and 263):

x′_i = (x_i − x_c) / d_io ,     y′_i = (y_i − y_c) / d_io

where (x_c, y_c) denotes the centroid coordinates. This normalization ensures invariance to camera placement distance and minor translational offsets.

C. Embedding Generation and L2 Normalization

The normalized 936-dimensional feature vector is projected to a compact 128-dimensional embedding via a learned linear projection matrix W ∈ ℝ^{128×936}, trained on an internal enrollment dataset using a contrastive learning objective. The embedding e is L2-normalized:

ê = e / ‖e‖₂

L2 normalization constrains all embeddings to the unit hypersphere, so cosine similarity between query ê_q and reference ê_r equals their dot product:

S(ê_q, ê_r) = ê_q · ê_r = Σᵢ₌₁¹²⁸ ê_q^(i) · ê_r^(i)

An identity decision is produced by comparing S against calibrated threshold τ: a probe is accepted as identity k* = arg max_k S(ê_q, ê_k) if S(ê_q, ê_k*) ≥ τ. The operational threshold τ = 0.85 is derived from ROC analysis described in Section III-D.

D. Threshold Calibration Methodology

Threshold calibration was performed via leave-one-out cross-validation. For each candidate τ ∈ {0.70, 0.72, …, 0.98}, FAR and FRR were computed across genuine and impostor probe pairs. The Equal Error Rate (EER) operating point (FAR ≈ FRR) was identified at τ = 0.83. The operational threshold τ = 0.85 was selected to bias toward lower FAR, reflecting the institutional priority of preventing proxy attendance. Table II summarizes selected calibration results.

TABLE II.  Threshold Calibration Results  (N = 70 Participants)

E. CSV Logging and Duplicate Prevention

Attendance records are written to structured CSV files with schema: [StudentID, FullName, Date, TimeIn, SessionID, ConfidenceScore]. Duplicate prevention is implemented via an in-memory session registry tracking verified identities within a configurable 60-minute window; each identity is logged at most once per session. Fig. 2 presents the complete processing flowchart, and Algorithm 1 provides the pseudocode for the verification and logging loop.

Fig. 2 – Verification and Logging Processing Flowchart

Algorithm 1 – Core Verification and Logging Pseudocode

EXPERIMENTAL EVALUATION

A. Testing Methodology

Evaluation was conducted across N = 70 university students (42 male, 28 female; age 18–26). Informed consent was obtained prior to enrollment. Each participant contributed five enrollment images under studio-grade illumination (3200 lux, CRI > 95) at 0°, ±15° horizontal pose. Three environments were assessed: (i) Controlled – uniform 2800 lux overhead lighting, background-free, subject distance 60–80 cm; (ii) Variable Illumination – 120–4200 lux range with mixed light sources; (iii) Real-World Classroom – functioning 35-seat lecture room with projector glare, distances 80–180 cm, and natural daylight.

Evaluation comprised 2,100 genuine probe trials and 6,300 impostor trials (cross-participant pairings). All experiments ran on a Dell Inspiron 15 laptop: Intel Core i7-1165G7 (2.8/4.7 GHz), 16 GB DDR4 RAM, no GPU, Ubuntu 22.04 LTS, Python 3.10.

B. Recognition Performance

Table III presents recognition metrics across all evaluation environments. Controlled conditions yield 98.6% accuracy, establishing a performance ceiling. The real-world classroom environment—most representative of operational deployment—achieves 96.7% accuracy with FAR = 1.1%. Fig. 3 illustrates accuracy degradation below 400 lux, identifying supplementary lighting as the highest-impact infrastructure improvement.

TABLE III.  Recognition Performance Across Evaluation Environments (N = 70, τ = 0.85)

Fig. 3 – Recognition Accuracy vs. Ambient Illuminance (lux) with 95% Confidence Intervals

C. Latency Analysis

Median per-frame inference latency of 38 ms corresponds to approximately 26 fps, consistent with smooth real-time display. Fig. 4 presents the latency breakdown and environment comparison. MediaPipe landmark detection accounts for 71% of total inference time (27.0 ms), indicating that future optimization efforts should prioritize landmark throughput, for example through model quantization or parallel frame prefetching.

Fig. 4 – Inference Latency Breakdown (left) and Latency by Evaluation Environment (right)

D. Scalability Analysis

Fig. 5 shows recognition accuracy and database lookup latency as enrollment size is varied from 10 to 500 identities (augmented via geometric transformations from the base dataset). Accuracy remains above 97% for enrollment sizes up to 200 identities, declining modestly to 96.1% at 500. Lookup latency scales linearly at 0.31 ms per 100 enrolled identities. For deployments exceeding 300 students, approximate nearest-neighbor (ANN) indexing is recommended.

Fig. 5 – Scalability: Recognition Accuracy and DB Lookup Latency vs. Enrollment Size

E. Comparative Baseline Analysis

Table IV presents cost-normalized performance against representative baselines. The proposed system achieves accuracy comparable to fingerprint biometric systems at significantly lower hardware cost and offers near cloud-level accuracy without GPU or connectivity requirements.

TABLE IV.  COST-NORMALIZED PERFORMANCE COMPARISON

† RFID accuracy denotes token-read rate; effective identity-verification accuracy is lower due to card-sharing.

F. Statistical Validation

Statistical significance of the accuracy differential versus RFID baseline was assessed via two-proportion z-test (N = 2100 trials per condition): z = 7.83 (p < 0.001), confirming significance at the 0.1% level. Accuracy in the real-world environment is reported as 96.7% ± 0.8% at the 95% confidence level (Wilson score interval). ROC curves for all three environments are presented in Fig. 6.

Fig. 6 – ROC Curves for All Evaluation Environments with Operating Threshold (τ = 0.85) and EER Point Annotated