Dynamic Age and Gender Prediction in Videos: A Real-Time Approach with PAIV

Abstract

Facial video stream age and gender recognition is becoming immensely significant in terms of its use in surveillance and content moderation, user profiling and secure digital onboarding. In this paper, PAIV (a real-time system that integrates convolutional neural networks, VGGFace-based feature extraction, Haar cascade face detection and Kanade-Lucas-Tomasi tracking to jointly predict age and gender on a live video at 30 frames per second) is presented. The system is trained and evaluated on large public datasets and on additional real-world video scenarios with challenging illumination, pose variation, occlusion, and multiple faces. Experimental results show a gender classification accuracy of 95% and a mean absolute error of approximately 5 years for age estimation while maintaining real-time throughput. These findings demonstrate that integrating multi-output deep models with efficient face tracking enables accurate, scalable demographic analysis from video, and highlight remaining challenges related to low-light conditions, low-resolution inputs, and fine-grained age prediction that motivate future enhancements.