Multi-camera face detection and recognition in an unconstrained environment

Lee, Kian Huang (2023) Multi-camera face detection and recognition in an unconstrained environment. Final Year Project, UTAR.

Preview

PDF Download (1922Kb) | Preview

Abstract

Multi-camera face detection and recognition is an Artificial Intelligence (AI) based technology that leverages multiple cameras placed at different locations to detect and recognize human faces in real-world conditions accurately. While face detection and recognition technologies have exhibited high accuracy rates in controlled conditions, recognizing individuals in open environments remains challenging due to factors such as changes in illumination, movement, and occlusion. In this project, the multi-camera face detection and recognition is developed in unconstrained environment setting. The multi-camera solution can overcome these challenges by capturing images of individuals from different angles and lighting conditions, thus providing a more comprehensive view of the monitored area. The pipeline in this project consists of three main parts – face detection, face recognition, single and multi-camera tracking. A series of models training is done with the open-source dataset to build a robust pipeline, and finally, the pipeline adopted trained YOLOv5n for the face detection model with mean Average Precision (mAP) of 0.495. The system also adopted the SphereFace SFNet20 model with an accuracy of 82.05% and a higher inference rate as compared to SFNet64 for face recognition. These models are then fed into DeepSORT for multi-camera tracking. Our dataset has been applied to the pipeline and shown ideal outcomes with objectives achieved. The solution feasibility is demonstrated via prototype implementation.

Actions (login required)