Successful Facial Surveillance CaseBy: John Honovich, Published on Jan 07, 2013
Real time facial surveillance is hard to do - accuracy issues, false alerts, cost and deployment complexity all conspire against it. This is why, more than a decade after deployments started they still remain remarkably uncommon. However, one application - self exclusion list - offers great promise and at least one practical example. In this note, we examine what key features make it work where most others have failed.
Self Exclusions List
Certain municipalities establish self exclusion lists to casinos for people who have gambling addictions. The goal is to help those people stay away from casinos and, if they go, block their entry.
The challenge for casinos is that they have huge numbers of visitors plus thousands who are on the self exclusion list. This makes it very hard for even trained personnel to block those on the list.
Recently, Ontario implemented a facial surveillance system to solve this. From speaking with the manufacturer, a few logistical elements greatly help the solution that are not common in other use cases:
Subject Actively Enrolled
Unlike criminals, people who are on the self exclusion list voluntarily enroll into the system as such:
This greatly improves accuracy as the system has multiple, head on, high quality images. By contrast, facial surveillance systems often depend on grainy, outdated or bad angle images.
The manufacturer says that these casinos have few choke point entrances. This makes it far less expensive to deploy cameras / infrastructure. By contrast, many retailers and public facilities have many egresses, driving costs up significantly and creating issues when people walk in at angles to the camera.
No Need to Catch Everyone
Since the harm of having a self-excluded person enter the casino is modest, especially compared to a criminal, the system does not need to catch everyone. Even if it identifies some persons on this list, it can be viewed as successful. This allows tuning to reduce operational issues.
False Positive Minimized
Specifically, the manufacturer acknowledges that the systems are tuned for lower sensitivity, reducing the number of false positives that frequently drive operators crazy. Facial recognition sensitivity is almost always adjustable:
- High sensitivity ensures the highest percentage of subjects on the watch list are detected but increases the number of non subjects falsely matched
- Low sensitivity reduces the percentage of subjects on the watch list detected but decreases the number of false alarms triggered (i.e., non subjects matched)
Because of the 'adversary' being monitored, the system is tuned to help avoid one of the most common sources of system failure.
While this is not your typical security application, it does highlight a number of common operational issues and how they can be avoided in certain use cases.