An Intelligent CCTV-Based Face Recognition System for Real-Time Identification in Crowded Public Environments

Abstract

The rapid expansion of surveillance infrastructures across urban environments has significantly increased the volume of video data generated in real time. Manual monitoring of such large-scale surveillance networks is inefficient and prone to human error, particularly in high-density public environments characterized by crowd congestion, pose variations, illumination changes, and motion blur. Traditional face recognition approaches based on handcrafted features lack robustness under such unconstrained conditions. This paper presents a CPU-optimized deep learning framework for real-time face recognition-based criminal detection designed specifically for high-density surveillance scenarios. The proposed system integrates Multi-task Cascaded Convolutional Networks (MTCNN) for accurate face detection and alignment, followed by a custom-trained TensorFlow embedding model for identity recognition using similarity-based threshold matching. The framework is trained on multi-angle real-world datasets and deployed under CPU-constrained environments to ensure practical feasibility. Experimental evaluation under controlled testing conditions achieved an average recognition accuracy of approximately 95%. The proposed architecture emphasizes computational efficiency, scalability, and secure biometric handling, making it suitable for distributed public surveillance systems.