Machine and Deep Learning Based CCTV Surveillance Using FaceNet, MTCNN, and Haar Cascade for Enhanced Security

As the population continues to grow, ensuring the safety of individuals and places from theft or unauthorized access is becoming increasingly challenging. However, traditional CCTV surveillance systems have limitations, such as manual monitoring, which leads to inaccurate results and requires high computational power. Considering these challenges, this paper proposes a hybrid approach that combines the efficiency of Haar Cascade with the precise face localization of Multi-Task Cascaded Convolutional Neural Networks (MTCNN) and the accurate face recognition of FaceNet. The system processes CCTV footage in real-time by converting it into frames, detecting faces using Haar Cascade, generating facial embeddings using FaceNet, and classifying them using deep learning and machine learning classifiers. The system achieves 99.3% accuracy using Vision Transformers (ViT) on the custom dataset, while the Support Vector Machine (SVM) reached only 98.5%. The inference time of SVM is 12.3ms, and ViT is 20.2ms. SVM achieves efficiency in real-time. A lightweight Flaskbased graphical interface facilitates real-time monitoring. The framework offers a scalable, cost-effective solution for criminal monitoring and missing person identification with minimal human intervention. This work marks a significant advancement in public security and law enforcement infrastructure by providing an efficient alternative to conventional CCTV systems, with potential applications in congested public areas and edge-device deployments.