University of the Western Cape Case Study: From Existing CCTV to AI Video Intelligence

The University of the Western Cape (UWC) operates a complex security environment serving a community of more than 23,000 students and approximately 2,400 academic, professional and support staff. Campus Protection Services is responsible for functions including access control, CCTV monitoring, crime prevention and emergency response across the University’s facilities. Against this backdrop, UWC evaluated AVA-X as part of its ongoing assessment of technology for its security operations and new control-room environment.

The objective of the pilot was to determine whether facial recognition could operate effectively in a live university setting using existing camera infrastructure, while supporting potential applications such as monitoring staff entrances and exits, logging authorised access and identifying individuals who are no longer permitted to enter selected areas. The wider use case is intended to support an access environment involving approximately 4,000 people, a defined subset of the wider campus community, without requiring every individual to stop and present a physical badge at monitored entrances.

A live deployment, not a laboratory test

AVA-X was deployed across three entrance video streams on UWC’s main campus. Rather than installing an entirely new camera estate, the system received video feeds using RTSP over a secure connection. The deployment required no replacement of UWC’s existing surveillance infrastructure.

Following an initial 30-minute technical session between the UWC and AVA-X teams, the three streams were connected and operational within a further approximately 90 minutes. From the start of the technical session to live operation, the deployment took approximately two hours.

This rapid implementation was enabled by AVA-X’s flexible architecture and the use of a secure Tailscale network connection, allowing the platform to be introduced alongside UWC’s existing systems without an extensive network or camera reconfiguration project. The pilot ran across five consecutive calendar days, from Thursday through Monday, covering three working days (Thursday, Friday and Monday) and the intervening weekend.

Sovereign AI processing inside South Africa

For the UWC pilot, AI inference was provided through Gigabrain’s infrastructure in Johannesburg. AVA-X has a strategic relationship with Gigabrain around sovereign compute and AI infrastructure in Africa. For the UWC deployment, this allowed the Swiss-developed AVA-X technology to perform its AI processing within South Africa rather than relying on an overseas cloud environment.

The architecture demonstrated how advanced video intelligence can be deployed while retaining AI workloads within the country in which the system is operating. Despite inference being performed remotely in Johannesburg, there was no operationally perceptible delay in identification. During the deployment, AVA-X recognition inference operated in under 30 milliseconds. An on-premise deployment remains available for organisations that require processing to take place directly at their own facility.

Facial recognition under real-world conditions

The purpose of the UWC deployment was not simply to determine whether AVA-X could identify a clear, front-facing image under controlled conditions. The technology was evaluated against the type of imagery and environmental conditions encountered in an operational campus-security environment.

AVA-X demonstrated 98.2% facial-recognition accuracy during the live UWC deployment. This field result was achieved on operational video rather than controlled benchmark imagery, including recognition events involving partial facial obstruction, sunglasses and interference from physical structures such as the rolling steel entrance doors at the monitored access points.

The system also demonstrated its ability to work with low-resolution facial imagery. AVA-X’s recognition technology can create and compare facial vectors using facial image crops as small as approximately 42 × 42 pixels, enabling recognition even where the source imagery available to the system is considerably below ideal portrait quality.

The UWC environment also provided an important opportunity to demonstrate the technology across a diverse population and a wide range of skin tones in a real African operational environment. Combined with rapid inference and integration into the existing video streams, the pilot demonstrated that AVA-X could perform facial recognition without requiring ideal laboratory conditions or purpose-built camera infrastructure.

Methodology and benchmark context

AVA-X converts each detected face into a feature vector and treats two faces as the same person when the L2 distance between their vectors falls below a calibrated threshold, balancing false matches (FMR) against false non-matches (FNMR).

For benchmark context, the model achieves 99.82% verification accuracy on LFW and 99.87% on CFP-FF, as well as 98.31% on CFP-FP and 98.25% on AgeDB-30, which introduce greater pose and age variation.

These benchmark figures are not directly comparable with the 98.2% UWC field result. The benchmark tests evaluate controlled verification tasks, whereas the UWC deployment involved live video with moving subjects, occlusion and low-resolution facial crops. The 98.2% figure should therefore be read as real-world deployment performance rather than as the model’s laboratory accuracy ceiling. The pilot did not produce error rates broken down by demographic group.

Independent evaluations such as the NIST Face Recognition Vendor Test show that error rates can vary across demographic groups, and common benchmarks under-represent African faces. Structured, POPIA-compliant field evaluations – for which research universities such as UWC are well placed – would help close this gap, and AVA-X’s in-house model development allows findings to feed directly into training.

From recognition to intelligent access management

The pilot focused primarily on facial recognition, but the wider objective was to demonstrate how existing surveillance infrastructure can become part of a more intelligent security and access-control workflow.

Potential applications discussed during the deployment included identifying individuals placed on an authorised security watchlist, recognising people whose permission to access a facility has been withdrawn, and identifying former employees or other individuals who should no longer have access to particular premises.

The architecture also creates opportunities for integration with existing access-control and security systems. Instead of requiring every authorised individual to stop, locate a card and physically scan it, facial recognition could support identity verification and automatically record entrance and exit events as people move naturally through monitored access points. This provides a potential path towards frictionless access management while allowing organisations to continue using the security infrastructure in which they have already invested.

Privacy-conscious biometric processing

The UWC pilot was conducted using controlled, authorised participants who were aware of the deployment. The number and identities of those participants are not disclosed.

AVA-X does not rely on retaining a conventional facial photograph for ongoing recognition. During enrolment, facial characteristics are converted into a mathematical representation known as a facial vector or embedding. This vector is subsequently used by the recognition engine to compare identities. Because facial vectors still represent biometric information, they are treated as sensitive information rather than being regarded as anonymous simply because a conventional photograph is not being used for each comparison.

For the UWC deployment, AI processing took place within the South African sovereign-cloud environment provided through Gigabrain’s Johannesburg infrastructure. Access to the pilot environment remained controlled by the participating teams. At the conclusion of the pilot, the biometric information created specifically for the evaluation was removed from the system. This approach demonstrated how facial-recognition technology can be deployed with defined enrolment, controlled access to biometric information and a clear end-of-pilot data lifecycle.

Adding intelligence without replacing infrastructure

One of the most significant outcomes of the UWC deployment was the speed with which AVA-X could be introduced into an existing security environment with different camera types and vendors. Organisations frequently have substantial investments in CCTV, networking and access-control systems. Requiring those organisations to replace cameras before introducing AI can dramatically increase the cost, complexity and implementation time of a project.

The UWC pilot took a different approach. Three video streams were securely connected and made available to the AVA-X recognition platform in approximately two hours. There was no rip-and-replace project and no requirement to rebuild the University’s surveillance estate. The intelligence layer was added to the infrastructure already available.

This interoperability has implications beyond higher education. The same deployment model can be applied to corporate campuses, industrial facilities, transportation environments, healthcare facilities, public infrastructure and other organisations with substantial existing CCTV investments.

Extending the value of existing CCTV infrastructure

Because AVA-X works on standard video streams, further AI models can run on the same RTSP feeds and sovereign compute. The following capabilities were not part of the UWC pilot.

• Weapon detection. Real-time alerts for visible firearms and other weapons help control-room operators focus on critical incidents.
• Number-plate recognition and driver-vehicle association. ANPR logs vehicle access and, combined with facial recognition, can verify that an authorised vehicle is driven by an authorised person.

Each model adds value to existing infrastructure, subject to its own evaluation, defined purpose and compliance with data-protection law.

Built for the environment in which it operates

For AVA-X, the UWC pilot represents an important African deployment because the technology was evaluated where it is ultimately expected to operate: against live video streams, existing infrastructure, ordinary movement, imperfect image conditions and a diverse population.

UWC’s security environment provides a particularly relevant setting for such an evaluation. Its security operations include CCTV monitoring, access control and broader campus-protection responsibilities, while its Director of Risk and Compliance, Bentley Ebrahim Jeftha, also serves as Deputy President of the Campus Protection Society of Southern Africa.

The pilot therefore represented more than a technology demonstration. It tested whether AVA-X could be deployed quickly, integrate with an existing environment, perform recognition under real operating conditions and use sovereign African compute infrastructure while maintaining the responsiveness required for security applications.

For AVA-X, the significance of the deployment is straightforward: advanced facial-recognition technology does not need ideal laboratory conditions, brand-new camera infrastructure or overseas AI infrastructure to perform effectively in Africa. The UWC deployment demonstrated that it can operate in the environment in which it is intended to be used. Discussions regarding potential future applications of the technology at UWC are ongoing.

Three principles demonstrated by the deployment

1. Use the cameras organisations already have.
2. Keep sensitive AI workloads within the jurisdiction and infrastructure appropriate to the customer.
3. Deliver recognition quickly and accurately enough to support real-world security operations.

University of the Western Cape comment

“The setup process was straightforward and effortless. The system was easily integrated with the University’s existing cameras without requiring major changes to the existing infrastructure. The facial recognition functionality performed well during the POC, with accurate identification results and a very low incidence of false readings.”

Bentley Ebrahim Jeftha
Director of Risk and Compliance
University of the Western Cape

Background sources

• University of the Western Cape – UWC at a Glance
• University of the Western Cape – Risk and Compliance / Campus Protection Services
• University of the Western Cape – Bentley Ebrahim Jeftha elected Deputy President of CAMPROSA
• Gigabrain Africa – Sovereign AI infrastructure
• AVA-X – Face Recognition Verification Test (white paper, February 2025)
• Huang, G. B., Ramesh, M., Berg, T. & Learned-Miller, E. (2007) – Labeled Faces in the Wild. Technical Report 07-49, University of Massachusetts, Amherst
• Sengupta, S. et al. (2016) – Frontal to profile face verification in the wild. IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1–9
• Moschoglou, S. et al. (2017) – AgeDB: The first manually collected, in-the-wild age database. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 1997–2005
• Grother, P., Ngan, M. & Hanaoka, K. (2019) – Face Recognition Vendor Test (FRVT) Part 3: Demographic Effects. NISTIR 8280, National Institute of Standards and Technology
• Protection of Personal Information Act 4 of 2013 (POPIA), Republic of South Africa. Government Gazette No. 37067, 26 November 2013

About AVA-X

The AVA-X group operates across three entities: AVA-X AG (Switzerland), where the core technology originated; SAFEZA AVA-X (South Africa), the operational proving ground; and AVA-X, Inc. (United States), the contracting and support entity for the American market.

Explore our AVA-X products, or get in touch to discuss campus, corporate, and critical-infrastructure deployments.