MTS makes a building's CCTV cameras work as one. It detects people at the edge, re-identifies them across views, and maps them to a floor plan.

MTS (Multi-camera Tracking System) started from a simple question. Can a building's CCTV cameras work together to understand how people move through a space? It was my team's build for the second challenge at the Apple Developer Institute for AIML, the visual-data counterpart to the audio challenge that produced Pezen. Where Pezen was about making sense of sound, this one was about turning raw video into something meaningful about the real world. There were three of us, Khresna, Davin, and me. We drew inspiration from a hockey-tracking project called HockeyNet, and we had less than three weeks. What we came out with is a proof of concept, not a finished product, and I will be upfront about its rough edges, but it does something I am genuinely proud of.

A single camera can tell you a lot, and our PPE system SCM already runs detection across several cameras at once. What it cannot do is realize that the person leaving camera one is the same person entering camera two. MTS is built around exactly that. It has three jobs. First, detect every person in every camera. Second, re-identify them, so one visitor walking past four cameras is counted as one visitor and not four. Third, place all of it onto a single floor plan, so instead of four separate video feeds you get one top-down view of where everyone actually is. Detection, identity, and location. Together they turn a wall of CCTV monitors into an understanding of how a space is used.

The first job runs where the cameras are. Just like SCM, each camera is paired with a Raspberry Pi 5 and a Hailo-8L A low-power AI accelerator chip for running vision models on edge devicesaccelerator, and we fine-tuned a YOLO11 A fast, modern object-detection model, used here to find people in each framemodel to run on that hardware in real time. Doing detection on the edge means only lightweight results, not raw video, travel onward, which keeps the system fast and keeps most of the footage on-site. Each edge device watches its own camera and sends a compact stream of the people it sees to a central processor.

The central processor is where the identities come together. It takes the detections from every edge device and runs StrongSORT, A tracking algorithm that follows detected people frame to frame and links them across cameraswhich tracks each person frame to frame and tries to re-identify them across different cameras, so the same visitor keeps the same identity as they move through the building. This is the hard part, and I want to be honest about it. Telling two similarly dressed strangers apart, across cameras with different angles and lighting, is genuinely difficult, and in under three weeks we did not get the Re-ID Re-identification, recognizing that a person seen by one camera is the same person seen by anotheraccuracy to where I would want it. It works and it demonstrates the idea, but it is the first thing I would sharpen with more time.

The last job is the one that makes it click visually. Each camera sees the floor at an angle, so we used homography, A geometric mapping that warps a camera's angled view onto a flat top-down plana classical geometric mapping, to warp every camera's view onto a shared floor plan. The result is a virtual top-down map of the space, where a person spotted by an angled camera in the corner shows up as a dot in the right place on the plan, and observations from different cameras merge into one picture. That is what turns tracking into insight. You can read traffic patterns, see which areas are busy, and follow how people flow, all from cameras that were already on the wall.

MTS is nowhere near production, and I would not pretend otherwise. It is a proof of concept built under a tight clock, with Re-ID accuracy as its clear weak point. But it stacks three computer-vision techniques on top of each other, edge detection, cross-camera Re-ID, and homography floor mapping, and makes them work together end to end, which is exactly what the challenge was asking for. The applications are easy to imagine, from retail analytics to crowd safety to how people move through a hospital or a station. More than anything, it taught the three of us how to deploy AI at the edge, track people across cameras, and turn raw video into something you can act on. It is the concept I am most excited to keep pushing further.