Sentinel is an intelligent multi-camera surveillance system that automatically tracks vehicles across an entire city's camera network, without a human having to watch a single second of footage.
Imagine someone walks into a police station to report their vehicle stolen. The officer on duty turns to the Jamaica Eye camera network — the city's best shot at finding it. But there is no automated system. Instead, an officer must manually pull up feed after feed, scrubbing through hours of footage from dozens of cameras, hoping the vehicle appears in frame.
It is slow. It is exhausting. And in the time it takes, the vehicle is long gone. This is the reality facing law enforcement today. Sentinel was built to change that entirely.
Sentinel's pipeline runs continuously: watching, learning, and reasoning. When you need to find a vehicle, the answer is already waiting.
This is Sentinel's most distinctive original contribution. With dozens of cameras feeding footage simultaneously, Sentinel cannot process them all at once — so it built a sophisticated priority-based scheduling system to decide which camera gets processing power at any moment.
Every camera has a calculated priority score based on its road network importance, how recently it was processed, and whether an active search has flagged it as relevant. When you search for a vehicle, Sentinel automatically boosts the priority of nearby cameras, directing compute resources exactly where they are most needed. A built-in anti-starvation mechanism guarantees that no camera is ever permanently ignored.
Once a camera is scheduled for processing, each frame is passed through Sentinel's detection layer. This layer uses YOLOv8 (a high-speed object detection tool) and OSNet (an embedding tool) as components to find, classify, and fingerprint every vehicle in the frame. These are tools Sentinel integrates — they are not what Sentinel is.
On standard hardware, this layer processes each frame in 300 to 500 milliseconds — analysing, classifying, and fingerprinting vehicles far faster than any human eye, continuously, across every scheduled camera without pause.
Sentinel uses PaddleOCR (a plate-reading tool) to read the licence plate from footage. To ensure accuracy, it reads the plate multiple times and only stores the result once at least 3 readings agree. In the example shown, all 3 readings matched, giving 95.97% confidence in plate "3324LG".
But here's what makes Sentinel special: even when the plate isn't visible, it still recognises the vehicle by its visual "fingerprint," using deep learning identity embeddings that capture its shape, colour, and distinguishing features.
Sentinel builds a complete timeline of every confirmed sighting (which camera, at what time) and displays it on a live map with numbered stops showing the vehicle's exact route.
When there's a gap between cameras, Sentinel flags it honestly so operators always know what's confirmed versus inferred.
Sentinel uses a Markov Chain, a mathematical model that analyses sequences of events to calculate probabilities, to predict the most likely cameras a vehicle will appear at next. Based on the route it has already taken, Sentinel calculates and ranks predictions in under a second — giving officers a head start rather than a guess.
These predictions get smarter over time, learning from every vehicle that passes through the network. The more it observes, the more accurate it becomes.
Every minute spent searching footage is a minute not spent on the ground. Sentinel closes that gap.
What currently takes officers hours of manual footage review can be completed in minutes. An officer types in a licence plate and Sentinel returns a full reconstructed route across all cameras automatically, so they can act immediately rather than search endlessly.
Traditional surveillance only helps after the fact. Sentinel's Markov Chain prediction tells officers where a vehicle is likely heading next, turning a reactive search into a proactive intercept. Law enforcement gains the upper hand rather than playing catch-up.
Sentinel requires no expensive commercial licences, no cloud subscriptions, and no specialised hardware. It is designed to run on the infrastructure law enforcement agencies already have, making it immediately deployable without additional budget.
Sentinel is the system we built. To bring it to life, we selected and integrated the best available open-source tools, all running on standard hardware with no expensive cloud services required.
We used YOLOv8, a high-speed object detection library, as the tool that spots and classifies every vehicle in every camera frame in real time.
We built a deep learning identity system that gives each vehicle a unique mathematical "fingerprint," allowing Sentinel to recognise it even when the licence plate isn't visible.
We integrated PaddleOCR, an optical character recognition library, as the tool responsible for reading licence plates from footage at speed and cross-checking readings for accuracy.
We designed a Markov Chain model, trained on observed traffic patterns, that predicts the probability of a vehicle appearing at each connected camera based on its current route.
We built a formally proven scheduling algorithm that ensures every camera is processed fairly and boosts the cameras most relevant to any active search.
We designed a real-time web interface that brings everything together: live feeds, interactive maps, route timelines, and persistent investigations in one place.
Sentinel was built specifically for the operational realities facing law enforcement agencies in Jamaica and across the Caribbean. Limited cameras, constrained budgets, and no access to expensive commercial surveillance platforms — these are not obstacles Sentinel works around. They are the conditions it was designed for.
It runs entirely on open-source tools, requires no cloud subscriptions, and operates on standard hardware already available to agencies. As Jamaica's camera network grows and image quality improves over time, Sentinel scales with it — no rebuild, no re-procurement, no added cost.
The goal is simple: give law enforcement the intelligence layer they have never had, starting today, with what they already have.
Final-year Computer Science students at The University of the West Indies, Mona, passionate about using technology to solve real problems close to home.
Department of Computing · The University of the West Indies, Mona · COMP 3901 Capstone, April 2026
A complete, open-source, AI-powered vehicle surveillance system, built from scratch by four undergraduate students with formal mathematical proofs and a live working demonstration.