Recording video for DataFromSky

This page details all possible aspects of recording video for DataFromSky. In case you think something is missing here, or have further questions, do not hesitate to contact us and ask!

1. Legal and regulatory concerns

Even if DataFromSky is based in Czech Republic (EU), the locations where traffic videos for DataFromSky are recorded are scattered around the world. Therefore, DataFromSky depends primarily on their Service Partners and subcontractors to observe laws and regulations at the location of recording the traffic video. These typically include flight height, distance from urban areas and roads, tethering, notices to airmen etc. and will often conflict with other requirements outlined below.

unidentifiable pixelated cars

1.1 Personal information

DataFromSky analyzes detailed movement information of individual vehicles and aggregates selected characteristics to provide a statistical overview of traffic behaviour.

To our best knowledge, DataFromSky is incapable of identifying individual vehicles or their drivers and does not store any personal information. Vehicles are seen from top, at a distance and image resolution which prevents any hypothetical reading of license plates or identification of persons. (See illustration to the right.) Vehicles bearing distinctive marks, labels, unusual features etc. may be identified only when the vehicle’s appearance is matched to identification from some source other than the DataFromSky video.

2. Flight and camera considerations

The following is written with a rotary wing UAV in mind, but most of the guidelines apply to any method of data collection.

  1. An ideal video for analysis by DataFromSky is exactly top-down. Since that is almost never achievable, the angle under which vehicles are machine-recognizable is currently about 30°. This figure relates to angle between camera axis and ground – the angle for distant corners of field of view will always be greater, of course.
  2. DataFromSky performs some optical stabilization itself, so the normal rocking of UAV in wind is tolerable. Gimbals etc. can help further. The whole area to be analyzed should be always completely in the picture, though.
  3. The most important restriction for DataFromSky video is size of a vehicle in pixels. The minimum is 16 pixels, which depends on sensor resolution, flight height and lens. Generally, shooting the video from 100-150 meters (300-500 feet) provides a reasonable tradeoff between image sharpness and covered area.
  4. The camera should be aimed away from the sun, so that there are no flares and changes in exposure. Strong sunlight also causes reflections off vehicles – it is advisable to avoid it, as the reflections can interfere with stabilization of video. Broadly speaking, a cloud cover providing soft, omnidirectional ambient lighting is best.
  5. Weather conditions can directly impact visibility and thus quality of video. Usually, flying an UAV in such conditions which would be optically problematic is considered a hazard, so these conditions are to be avoided.
  6. Camera (UAV) position change during recording is possible, as long as all of the interesting area is covered. A fixed-wing UAV orbiting the area is also a possibility, as long as the other conditions are reasonably fulfilled (particularly occlusion).
  7. Tall buildings protruding into field of view can occlude roads and thus drastically decrease actual visibility. The same holds true for trees, sometimes grown next to the roads.

3. Camera calibration

In order to create a precise model of the target scene, DataFromSky must know how to warp the image from camera to a surface with the real shape. The first step in this process is correcting the optical distortion caused by the imaging system itself, so that the image corresponds with a projection onto a hypothetical flat surface.

There are several types of distortion incurred by the lens, at different magnitudes, all at once. Usually, the most pronounced distortion types in a commercially available camera will be replicable across all of the cameras of that model, so in some cases it will be enough to know the camera model. In case the camera does not fall into a class of devices with known distortion, though, a calibration must be performed.

Please be aware that this distortion is constant for any given camera only as long as the optical system is completely fixed! Any camera with adaptive focus cannot be calibrated, unless the distortion is known at a sufficient amount of states of the camera, and that information is somehow recorded along with the video.

Note: Change of image resolution requires recalculation of the distortion parameters, although not a whole new calibration.

3.1 Calibration procedure

In order to calibrate a camera, the following steps must be performed:

  1. Print a chess board test pattern on a clean sheet of paper, so that its squares remain square, and is not blurred by resizing.
  2. Measure its size.
  3. Attach it to a hard surface so that it is completely flat.
  4. Take the camera in question (and lenses if applicable) and position it so that it looks at the test pattern.
  5. Capture a short video of the test pattern with the camera, at the resolution to be used, while moving the camera so that there are pictures of test pattern in various parts of the field of vision. (See also an example of the video.)
  6. Provide the video to DataFromSky team for analysis.

A quantitative description of the distortion can then be found – and later corrected for.

points marked with letters in picture from video
gps coordinates corresponding with letters from picture

4. Georeference

To measure lengths and thus also speeds and acceleration as accurately as possible, DataFromSky needs some reference for size of ground features, to warp the image correctly. This information is supplied as a list of points in the picture and their coordinates.

There are several steps to success with georeference in DataFromSky. The first is to select spots that are always visible in the video, with no temporary occlusions, do not change, and are very well distinguishable by the software from their surrounding. This rules out eg. trees (swaying gently in the wind!), pedestrian crossings (people walk on them), large tiles (may be confused with neighbour) etc.

The second step is obtaining GPS coordinates for these points. You can walk to these points and write down readings from your portable GPS equipment. Of course, you want to verify that the distances between these points do actually work out to something related to reality! Measuring small distances with GPS is sometimes a very tricky task. If there are maps with a good aerial imagery coverage in the area, you can use these, as long as there are no obvious seams of the picture going through the analyzed area. Either way, GPS coordinates are what you want.

Alternatively, a local coordinate system can be employed for analysis instead of global. In that case, you will have to provide the points, and some meaningful measurements of the distances. This will require some triangulation, or clever use of right angles present in the area itself.

Finally: Finding the data for georeference is somewhat independent of the recording act itself. These data can be collected after the video comes into existence. Still, it must be done before any analysis can start.

5. Multiple video sources

Generally speaking, DataFromSky can merge video image streams from multiple moving sources. Thus, it is possible to “hot-swap” UAVs in a position over target area to provide continuous coverage, while a single UAV would not last long enough. The considerations for all of the UAVs are identical to a single UAV, as far as video recording is concerned.

An important requirement is some way to precisely synchronize the image streams. You can start by properly setting the clock of all cameras. Even more importantly, do follow by providing some visual cues in the part where videos overlap.

The actual merging of videos is a task requiring complicated manual handling of the inputs, though, so it is best to minimize the amount of image sources involved in the whole process.

6. Delivering the video to DataFromSky

You can upload video at