Design and Implementation of a Real-Time Face Recognition System Using OpenCV and LBPH

Abstract

This paper presents the design and implementation of a lightweight real-time face recognition system using OpenCV that combines Haar Cascade face detection with Local Binary Patterns Histograms (LBPH) recognition for CPU-only, low-cost deployments. Practical applications require fast identification with limited computational resources; therefore, this study investigates whether a Haar Cascade–LBPH pipeline can achieve strong recognition accuracy on a 50-subject indoor dataset using a confidence threshold of 50 or below and examines how performance degrades when facial yaw rotation exceeds 30° and occlusion exceeds 40%. The contribution lies in an implementation-grounded report that explicitly specifies the operational decision rule for confidence-thresholding and documents empirical boundary conditions for rotation and occlusion, which are frequently insufficiently reported for classical pipelines. The system is implemented in Python with OpenCV (including the cv2.face module) and organized into three modules: ROI-based face acquisition and deterministic labeling via an s1–s50 directory structure, LBPH model training with persistence, and real-time inference from a USB camera stream. Experiments show that the system achieves 89.7% recognition accuracy on 50 subjects under confidence ≤ 50, while accuracy declines for more ambiguous score ranges and degrades markedly beyond the identified rotation and occlusion limits. These results indicate that the proposed pipeline provides an interpretable and practical solution for controlled indoor scenarios, although conservative thresholding improves reliability at the cost of a narrower operating envelope under non-ideal capture conditions.