SCAN: Surveillance Camera Array Network for Enhanced Passenger Detection

Abstract

The automated detection of individuals within vehicles, with minimal to no human intervention, hold multifaceted implications in contemporary contexts. These applications span from aiding emergency responders and optimising transportation networks to facilitating automated crash response mechanisms and enforcing regulations concerning High Occupancy Vehicle and High Occupancy Toll lanes. In this paper, we introduce our camera system designed for passenger counting, leveraging five IP cameras equipped with a range of optical filters, employing image registration techniques, and integrating the YOLOv8 object detection model. The Surveillance Camera Array Network (SCAN) operates within the near-infrared domain of the electromagnetic spectrum in conjunction with the visible part. Four VIVOTEK IP cameras are outfitted with near-infrared, neutral density, polarising, and ultraviolet optical filters, while the final camera retains its stock lens. Our primary challenge lies in managing variable lighting conditions throughout the day. However, during nighttime, we achieve nearly perfect image capture of vehicles. To mitigate noise, glare, and other impediments, we initially apply camera calibration, image preprocessing, cropping, image registration, and finally, image fusion. Our findings demonstrate that our cost-effective SCAN system adeptly detects passengers in cars equipped with window tinting. The results obtained during testing conditions resulted in an 66% true positive rate, 8% false positive rate and 26% false negative rate within best dataset. Additionally, we provide created datasets displaying passengers inside Suzuki Vitara, Jaguar XF, and Honda CRV vehicles with various levels of window tinting, to facilitate future community endeavors in addressing this challenging task.

Keywords

• Camera calibration
• digital image processing
• HOV
• image fusion
• image registration
• machine learning