Traffic Monitor
Vehicle and pedestrian detection, incident alerting, and 3D reconstruction from fixed cameras.
The nine screens on the home page are the system running, uncut — one incident each. Every vehicle and every person in frame carries a detection box, a class and its own tracking ID, held as they cross a junction and leave it rather than re-assigned frame to frame.
When something goes wrong the system says so in its own words, on the frame: potential collision, sudden stop, abnormal trajectory, dangerous driving, wrong way, pedestrian on road. And they escalate — a person stepping into live lanes in Fresno raises an amber warning that turns red as they reach moving traffic, and a car crossing a roundabout badly collects four flags that end on ACCIDENT LIKELY.
Behind the cameras is an operator board: recurring congestion by hour, lane utilisation, peak activity, and pedestrian zones measured by footfall and average dwell time. A tally counts all twelve turning movements through an intersection and increments as each vehicle completes one, and a 3D reconstruction draws every tracked trajectory in the colour of the movement it belongs to.
Which junctions, for whom, at what scale — unconfirmed. Two of the nine feeds run on MOT17-04, a public tracking benchmark, rather than on a deployed camera.
Every layer