FaceTrack: Deep Learning-Based Person Detection, Recognition, and Timestamp Extraction in Surveillance Videos

Abstract

The widespread deployment of Closed Circuit Television (CCTV) cameras in public areas, schools, transport hubs and retail outlets has resulted in the constant production of vast amounts of surveillance data. It is time-consuming and inefficient to manually check these videos to see if a certain person is in the video, especially for long videos. In this paper, we introduce a lightweight computer vision framework, called FaceTrack, for automatic face detection, person recognition and entry/exit timestamp estimation of a target person from CCTV videos. The proposed framework is based on Haar Cascade classifier for face localization and cropping, and a fine-tuned MobileNetV2 network for extracting discriminative facial embeddings for identity recognition. To improve the reliability of recognition and reduce false detections, Euclidean distance-based similarity matching, adaptive thresholding and framewise verification are used. The model was trained and tested on a custom face dataset consisting of about 50-60 augmented images for each of five individuals, with data augmentation performed using image rotation, zooming, and horizontal flipping. Experimental evaluation achieved a recognition accuracy of 91.78%, with a precision of 92.62%, recall of 91.78%, and an F1-score of 91.69%. The proposed system not only identifies the target person but also automatically records the entry time, exit time, and total duration of the target individual's presence, making it suitable for intelligent surveillance and forensic investigation applications.