An autonomous vision and motion system for commercial vehicle capture.
The client identified commercial vehicles from the roadside using photographers with long lenses, panning by hand. We engineered the system that replaced them, covering detection, speed estimation, camera motion and capture, and delivered a working unit that now runs unattended.
The client collects identification data on commercial vehicles from the roadside in Singapore. The work was done by DSLR photographers: standing beside an expressway with long lenses, tracking each vehicle by hand as it passed, and capturing both the licence plate and the operator marking on the vehicle’s side within the same pass.
The method produced usable results but did not scale. A photographer holds the task for around three hours before accuracy falls away, and every hour of coverage costs an hour of someone’s attention. The client specified a unit that could be mounted at the roadside and left to run.
The defining requirement was timing. Two of the three photographs are taken from a single rotating axis that has to be square to the vehicle at the moment it passes, and the rate that axis must turn differs for every vehicle.
A person panning a camera matches the vehicle by eye. The system does it by measurement, continuously, as the vehicle approaches.
Three photographs have to be taken of the same vehicle while it travels past at expressway speed. One is straightforward: a fixed long lens on the licence plate, fired as the vehicle crosses a known line. The other two are not. They sit on a single rotating axis that has to be square to the side of the vehicle at the instant the vehicle is beside it, and the target on that side, a marking near the front wheel, is only a few hundred millimetres wide.
The required rotation rate is not a constant. It is a function of the vehicle’s speed and its distance from the unit, and neither is known until the vehicle is already inside the measured area. A vehicle in the near lane needs the axis turning faster than the same vehicle further out. Get the rate wrong and the marking leaves the frame. Start the rotation at the wrong moment and it never enters it. The environment is uncontrolled: a closely followed vehicle places a second object inside the same measured area, and reacting to the wrong one photographs the wrong vehicle.
The rotation also has to continue through the exposure. This is what a photographer does when following a moving subject: while the camera tracks at the subject’s angular rate, the subject stays sharp at a shutter speed that would otherwise blur it. Halting the axis for the exposure removes that benefit and returns the motion to the image, so capture occurs mid-travel and the axis continues past it before resetting. Every camera on the array is global shutter, so the whole frame is exposed at one instant and a vehicle at speed is recorded without geometric skew.
Hardware selection and the physical arrangement were the client’s. We engineered everything that makes the arrangement behave as an instrument: what it watches, when it decides, how fast it moves, and what it keeps.
Work done by hand at the roadside, now carried out by an unattended unit.