At a Glance:
- Computer vision AI systems are a subset of artificial intelligence and machine learning that focus on making sense of digital images, videos and other visual inputs.
- Biometrics, or measuring human body characteristics, is proving to be a key enabler of digital identity.
- Soodeh Farokhi, CTO, C2RO, draws a distinction between facial recognition technology that identifies individuals versus de-identification categorization technology.
- One strong driver for the adoption of computer vision enhanced with AI can be seen in the delivery of healthcare, especially during the pandemic.
Last year, the revelation that Cadillac Fairview, a Canadian commercial real estate company, had embedded inconspicuous cameras in 12 shopping malls across the country and used facial recognition technology to analyze the age and gender of shoppers caused an uproar.
Shoppers would have been none the wiser, except a software glitch tipped off passersby and it was revealed that demographic data had been collected without consent. After an investigation was launched, the federal privacy commissioner found that Cadillac Fairview had contravened privacy laws by failing to obtain meaningful consent.
Computer vision AI systems are a subset of artificial intelligence and machine learning that impacts how we perceive and interpret information. The field focuses on making sense of digital images, videos and other visual inputs using cameras, data and algorithms. It has a wide range of applications, from facial recognition password protection (banking) to defect detection on factory lines (manufacturing), real-time health (healthcare) and sports tracking (consumer-lifestyle). It’s a lucrative field that’s expected to reach $48.6 billion by 2022.
But not all computer vision systems are created equal. Some systems apply combinations of image processing; statistical pattern recognition; and machine learning, deep learning and convolutional neural networks to accomplish tasks, including object detection and tracking, image classification or image retrieval and forecasting.
These technologies may be paired with identity verification technologies or biometrics, which measure either facial characteristics, fingerprints, your iris, the patterns your veins make, the way that you walk or other human traits, such as your EKG readings or voice. (All of these characteristics are unique to every person.)
Measuring human body characteristics and calculating behavior is proving to be a key enabler of digital identity and mobile remote identity verification systems, according to the Biometrics Institute. It offers a high level of efficiency and assurance, particularly when dealing with identity and credentialing.
At the same time, the use of biometrics (especially facial recognition systems) fills many with uneasiness because it depends heavily on large-scale collection of personal data. The Biometrics Institute’s annual Industry Survey revealed that even though a majority of respondents (63%) agreed that the pandemic has accelerated the adoption of biometric solutions, with three quarters reporting that new solutions and technology will be critical in managing the current and future pandemics, a significant proportion (48%) said the use of biometrics is growing too rapidly for existing controls to be effective.
Privacy Rights & Bias
Privacy concerns are the main barrier to adoption for AI video analytics. Add to that the lack of opportunity to consent, proportionality and the false assumption of machine neutrality, and the scope is significantly narrowed. For example, MIT Media Lab’s Gender Shades Project, an exploration of how “inadvertent negligence” seeps into digitalized and automated systems, found notable data bias and skews across data sets for gender and skin tone in leading tech companies’ AI systems. (These studies did not look at commercially available facial recognition algorithms.)
Like the Cadillac Fairview scenario, C2RO’s AI video analysis software platform aims to understand the human behavior journey in a physical facility. In-store journeys provide data on store traffic, demographics, social context (Is the customer alone, in a group or with family?), dwell time and staff engagement.
Unlike the Cadillac Fairview scenario, C2RO uses the facility’s existing security cameras to connect people from camera to camera without using any biometric information, stressed Farokhi, who holds a Ph.D. in computer science. Computer vision AI systems can capture patterns of behavior without compromising the identity or privacy of the individual, and should not be directly equated with pure-play biometric systems. All of these systems can be beneficial within as long as they are compliant with data privacy, Farokhi said.
Since C2RO’s platform traces the journey of a shopper through a mall without using vmetric information, Farokhi draws the distinction between biometric technology that identifies individuals versus de-identification categorization technology (which pinpoints, for example, the age range and gender, and analyzes behavior). That data is delivered anonymized and aggregated to strip out any personally identifiable information, she said, so it’s important to keep in mind that “face classification” should not be conflated with “face recognition.”
Is Responsible AI Worth the Effort?
Rather than view privacy protection as an unnecessary deterrent, C2RO leverages data privacy as a value proposition. “C2RO uses aggregated data, so we don’t have any record of any individual in our system and we do not store any biometric information,” said Farokhi.
She would not disclose the classification parameters or cues behind her company’s patent-pending technology (“It’s the secret sauce,” she said) but noted that the strictest privacy considerations are respected and embedded in the design.
This approach stems from the “privacy by design” framework devised by Ann Cavoukian, a leading privacy expert and former information and privacy officer of Ontario. The objective is to embed privacy into the design and architecture of networked infrastructure and ensure security and privacy algorithms extend throughout the lifecycle of the data. The concept is emphasized in the General Data Protection Regulation (GDPR), the privacy legal framework that has been in force in the European Union since 2018. It imposes obligations and penalties on organizations anywhere that target or collect data related to EU citizens.
Farokhi said that C2RO has elected to adhere to the GDPR privacy measures even though the requirements are more stringent than Canada’s privacy regulations. Under Canada’s Personal Information Protection and Electronic Documents Act (PIPEDA), entities are not allowed to collect, use or disclose personal information without explicit consent from the individual. This provision makes taking pictures in malls, video recording and analyzing the data (even for a millisecond) a contentious issue that can be contested—as Cadillac Fairview learned.
Similarly, as new legislative schemes arise in the U.S., employers using biometric information—to monitor online access to sensitive data, provide system access, monitor productivity and ergonomics—will have the added burden to meet legal and regulatory compliance requirements. While the Illinois Biometric Information Privacy Act is the only biometrics legislation that provides a private right to action, Texas, Washington, California, New York and Arkansas are working to catch up with their own biometric statutes or by expanding existing laws to include biometric identifiers, noted The National Law Review.
The Right Tool for the Right Purpose
One strong driver for the advancement and adoption of computer vision AI systems can be seen in the delivery of healthcare and medicine, where it is used to increase the speed and accuracy of diagnoses and disease screening and in support of public health interventions and outbreak response.
Many companies extended their offerings during the pandemic. For example, Kogniz AssuranceAI, a California-based start-up that coupled direct facial scanning with live video input feeds, rolled out its vaccine management functionality, which allows companies to track vaccine participation rates, send reminder notifications and incentivize employees to get vaccinated.
Scaling AI video analytics technology for new industries is part of Farokhi’s growth strategy. Irrespective of the industry, she maintains that understanding people’s behavior in relation to others and analyzing the human-machine interactions are part of the digital transformation. During the pandemic C2RO extended its solutions to include safety features, such as facemask analysis and real-time occupancy management for social distancing. These features, she said, will help businesses reopen with safety in mind and encourage them to share statistics on occupancy or sanitization efforts.
Still, a recent WHO report shows that opportunities for AI video analytics are inextricably linked to the risks, including the unregulated collection of data, the ethics of using the data, cybersecurity and encoded biases. The report sets out some of these risks and provides guidance on the ethics and governance of AI in healthcare systems.
“Here in North America—even with a compliant technology—people seem to be more skeptical about adopting it in comparison with people in Europe,” Farokhi said. “The boundaries of data privacy regulations are not yet very well defined and not as strict as it is in Europe.”
Farokhi contends that best practices will ensure organizations operate responsibly and hold all stakeholders accountable. Appropriate use of video analytics solutions, she said, will allow organizations to proactively support the safety, optimize business operations and boost their sales performance, while ensuring compliance with health and government regulations.
