VigilantGantry - Access Control with AI and Video Analytics
VigilantGantry is a fully automated, contactless gantry system for temperature screening developed by GovTech’s Data Science and Artificial Intelligence Division (DSAID). Driven by AI and deep learning, it augments existing thermal scanners to improve the rate of contactless scanning, ease bottlenecks in long queues outside buildings and reduce manpower required for temperature screening measures.
As part of a nation-wide effort to prevent and contain COVID-19 transmission, government buildings and public venues like shopping malls and workplaces have implemented temperature screening and access control measures. With Singapore gradually opening up, commercial buildings expecting high human traffic are making use of thermal cameras to scan and deny entry to feverish visitors.
However, to maintain a high level of vigilance, this approach comes with some difficulties: conventional thermal cameras cannot properly measure the temperatures of visitors with occluding headgear or hair fringes, requiring human intervention to have the obstruction removed. Conventional systems also require a human operator to monitor the output measurements, possibly for up to six hours at a time.
How does VigilantGantry work?
It is intelligent - Like a smart home hub, VigilantGantry links the gantry unit with existing thermal camera and CCTV systems. Face recognition, forehead exposure and colour segmentation algorithms provide the intelligence required for proper screening. VigilantGantry highlights any high temperatures detected by the thermal camera, triggering a customisable alert and denying entry to the visitor, whereupon further checks can be performed.
- With the YoloV3 person detection model, face segmentation is only triggered when the visitor is positioned within a pre-defined region of interest in the CCTV optical camera. The thermal camera system is hence prevented from indiscriminately scanning and triggering false alerts from background heat signatures such as mobile phones.
- In colour segmentation, the HSV colour model picks out very distinct colours which correspond to high temperatures. Red patches in the thermal camera reading may indicate the presence of a fever. For greater accuracy in detecting fevers, the algorithm also takes the size of the red patch into account.
- Visitors wearing headgear will affect the accuracy of temperature screening. Deep learning models are used in face segmentation to quantify the amount of exposed skin on a detected face. When these cases are detected through the CCTV optical camera, VigilantGantry will instruct the gantry to deny entry and direct the visitor to take off their headgear.
It is adaptable and scalable - The DSAID team designed VigilantGantry to integrate with existing thermal scanners without the need to purchase new equipment. For building owners, this means lower implementation costs and manpower needs, adapting to their current processes and plug-and-play scalability according to their needs.
It can support contact tracing efforts - Besides measuring skin temperature and performing facial indexing, VigilantGantry can also capture location, date and time details to aid contact tracing when required. It can also store health and travel declaration data obtained via questionnaires.
It is open-sourced - VigilantGantry’s face segmentation algorithm has been published on GitHub for the private sector to scale and deploy to sites across Singapore. The algorithm is used to detect visitors wearing headgear or with long fringes which may hinder the effectiveness of current thermal scanners.
To read more about VigilantGantry, click here.
For a technical walkthrough of VigilantGantry’s development process, click here.
Last updated 11 November 2020