Face Recognition Software Development
INDUSTRY
AI
SERVICE
Computer Vision
CLIENT
R&D
R&D: Face Recognition Software Development
"Waverley Software R&D team developed a face recognition system using Computer Vision and Deep Learning techniques."
Waverley's internal R&D team explored how to build a real-time face recognition system that identifies people entering a room from live surveillance footage and responds automatically, for example with a personalized audio greeting.
The team combined a deep learning face-embedding model with computer vision preprocessing and a two-tier detection cascade to work around common surveillance-camera limitations: poor lighting, low contrast, and the processing load of running recognition on every frame.
This was a self-funded research initiative rather than a client engagement. It has not moved beyond the prototype stage, and no production deployment or business outcome has been documented to date. The value of the project lies in the computer vision and deep learning depth it built inside Waverley's team, capability the company now applies to client work in security, hospitality, retail, and smart-building contexts.
R&D
Discovery
This project did not originate from an external client request. It was undertaken by Waverley's own R&D group, which regularly invests engineering time in applied AI and computer vision research ahead of client demand, so the team has hands-on experience with a technology before a client asks for it.
Face recognition was a natural target for this kind of investment. Identity recognition from video is a recurring requirement across security, retail, and smart-building projects, and the R&D team wanted first-hand experience with the deep learning and computer vision techniques involved before bringing that expertise into client engagements.
Project Analysis
The team set out to detect a human face in footage from a standard surveillance camera, match it against a database of known identities, and trigger an action, such as playing a personalized greeting or unlocking a door, when a match was found.
Two technical constraints shaped the approach.
Footage quality: surveillance video is often low contrast, unevenly exposed, and shadowed in ways that make faces harder to isolate and match reliably. A pipeline built for real camera feeds, rather than curated datasets, needed to correct for this before attempting a match.
Compute load: running a high-accuracy face detector on every frame of continuous video would overload the recognition server. The team needed a cheap way to filter out frames with no faces before spending compute on the frames that mattered.
What We Delivered
Waverley's R&D team built the recognition pipeline in three parts: identity matching, video preprocessing, and a two-tier detection cascade.
The Solution Includes
Face embedding and matching: a deep learning model generates a face embedding, a numerical vector representation, for each known identity and stores it in a database. When a new face is detected, the system generates its embedding and compares it against the stored vectors to find a match.
Video preprocessing: computer vision algorithms and optical filters correct for the contrast, shadow, and exposure problems common in surveillance footage, cleaning up the frame before it reaches the detection model.
Two-tier detection cascade: a fast, lower-accuracy detector runs first, on every frame, to separate footage that plausibly contains a human face from footage that does not. Only frames that pass this first filter go to a slower, higher-accuracy detector for confirmation, keeping the compute-intensive matching step off frames that do not need it.
Action triggering: when the slower detector confirms a match against the identity database, the system fires a notification and an associated action, such as playing a personalized greeting or opening a door
Let's Build Something Great.
Tell us about your project — we'll find the right path forward.