project settings

Sets the length and speed units together for every figure below. Acceleration is reported in g either way.

examples

Pick a validation case to load its tracked points and settings and see where the points sit in the video. Editing the tracked points returns the page to your own data.

reference length

A fixed distance on the object measured along the direction connecting the two tracked features. Wheelbase is the usual choice. Speed estimates strongly depend on this value.

When the object comes toward the camera or goes away from it at a steep angle, the length looks short in the image, so each pixel of placement error matters more. Enter the uncertainty accordingly and use frames where both features are sharp.

The range of the length is the width of the window within which the true length could reasonably lie, in the same unit as the length. If a wheelbase entered as 4.79 could be anywhere from 4.74 to 4.84, enter 0.10. The length is taken as uniform across that window, centered on the value entered. Each simulated trial draws one length and uses it for every frame, so this uncertainty moves all the speeds of a trial up or down together and does not average away over more frames.

Fills the true length from the Canadian Vehicle Specifications set, served by the NHTSA vPIC API. Applies only when the two tracked features are the front and rear wheel centers on one side of the vehicle.

curved path

Optional. On a curve the two tracked features, on one side of the object, follow a path of smaller or larger radius than the object's center. This scales the result from the tracked side to the center, after Choi et al.

The correction covers the difference in radius only, and assumes a constant radius and no sideslip. In simulation it recovers the speed for a camera looking straight down. A camera at an angle adds an error on a curve that this does not remove: about 1% or less for a view from the side, 10% or more for a view along the path.

pixel uncertainty

The width in pixels of the range within which each feature could reasonably be sitting. If the wheel contact patch could be anywhere across about three pixels, enter 3. Errors are taken as uniform across that range and independent between the front and rear points.

Judge it on the frame that reads most clearly and say which one that was. The range is carried to the other frames in proportion to the feature length, which holds the uncertainty in real units constant across the run. Choose Every frame to apply one range throughout.

acceleration limit

Optional. Each simulated trial moves the tracked points within the pixel range, and some trials produce speeds that jump between neighboring intervals by more than the object could manage. With a limit set, a trial is kept only if no pair of neighboring intervals implies an acceleration or deceleration above it. Trials continue until 2,000 are kept or the time allowed runs out.

With the limit on, the best estimate is the median of the kept trials, not the value computed from the points as tracked, and the bands are those of the kept trials. The note under the result reports how many trials were needed and what was kept. Check "Show individual trials" on the speed chart to watch the kept trials arrive.

tracked points

Use this when the camera was not stationary, or when your footage is a video of a screen. Any movement of the camera itself, whether panning, drifting, or zooming, looks like the object moving and will throw off the speed. By tracking a few things that are standing still, the tool can work out what the camera did and take it back out.

Track at least six points on the road surface the vehicle is driving on, such as markings, cracks, patches or joints. Spread them around the vehicle's path in the frames you use (beyond it, beside it and nearer the camera) rather than in one group, and keep every point in every frame. This applies whether the camera turned, moved, or both. Four points is the minimum the tool accepts, but more is better.

Using anything above the road surface, such as signs, poles, trees or parked vehicles, is not recommended.

For a video of a screen whose original camera was stationary, the four screen corners can serve instead. If the original camera moved, track road points inside the picture on the screen, as above. If the camera did not move at all, uncheck the box.

After calculating, read the Robustness line under the result. Under 1 percent, no single point controls the answer. Above 3 percent, consider adjusting your reference points or adding more before relying on it.