Impact Experiments - Imaging

Experiments carried out in the open in 1989 and 1992 were recorded using two NAC E-10 high-speed cine cameras and a video camera. These arrangements were far from ideal, but were a good compromise given the objectives (large field of view, uninterrupted trajectories). We did not have access to a vacuum chamber large enough for these experiments so the cameras needed to be far enough back from the impact to avoid damage. No photographs are presented here but the geometry of the arrangement is discussed in the later section. These early experiments provided a basis for the 3D analysis and a simplified geometry which has helped in carrying out subsequent work.

In the 1996, 1997 and 1998 runs at NASA Ames Vertical Gun Range (AVGR) we used three Milliken high-speed cine cameras for accurately recording the experiments, plus a NAC high speed video system. Fig. 1 shows the complete impact chamber with two cine cameras mounted above the chamber, slung upside-down below their tripods, and the last cine camera to the side with the video camera (seen close-up in Fig. 2). The arrangement of the top cameras, upside-down and in hard-to-measure positions, was unavoidable but gave rise to some problems in determining the imaging geometry (explained below).

The cine film provides the most accurate and high-resolution imaging but is not suited to any quick-look analysis. The video system allows an immediate minimal analysis of the shot - this facilitates the measurement of projectile velocity where necessary and highlights any fundamental problems such as an inaccurate shot (e.g. oblique when we wanted surface-normal) or damaged projectile; unfortunately the ice projectiles often broke up before reaching the target chamber. The NAC video is controlled from a trolley which is not visible here - it was generally located in the next room for safety reasons.

Various approaches have been used to get the cine film into digital form. These are summarised in the table below:

If you have the funds, the 1996 option is recommended, but you will also need a fast video digitiser card - ideally digitising full frames on the fly - in order to move from video to digital format. Also, the resolution of conventional video is quite low and may be inadequate for your needs. These factors really depend upon your software.

Given digital images, from multiple cameras, it is straightforward to extract 3D velocity data from these films. You need to know the geometry of your camera arrangement; this is usually simple to measure on site but we had problems with the NASA Ames experimental instrumentation because the cameras were somewhat haphazardly arranged (essentially this was determined by the available windows in the vacuum tank) and did not share orientation. Directly measuring the camera positions and orientations was very difficult because of their location on top of the tank, and their oblique views through the plexiglass windows.

Without precise measurements of the camera positions, you can determine the geometry a posteriori from the images themselves if you know accurately the positions of at least 5 non-coplanar points which are visible in each view. For example, in our case, we used a reference cube made of steel rods, which gives 8 points of reference. By using a view correlation program you can extract the key parameters of your setup; these are:

You will probably also need to note

These are the parameters you need to extract 3D data from your two or more cameras. The geometry is summarised in this figure. If your cameras have the same look point and up vector, and their eye points lie in a plane, you can do this very easily by application of trigonometric rules; you are basically triangulating the fragments. If, on the other hand, your cameras are in some crazy positions like ours, it is more difficult.

The 3D location approach used currently is to determine a vector from each camera to each fragment, then to determine the closest approaches for each of these vectors (they generally will not cross exactly). This gives you a convenient 3D position for each fragment in each frame. This approach is nice because it also gives you tangible error estimates in the form of the minimum distance between the vectors, and since the fragment position is calculated from scratch in each frame or timestep, you have distribution which can be used to investigate the relative significance of random and systematic errors. In our experimantal analysis, where these vector errors are typically calculated for some 50-100 points, the mean 3D position error is of the order of 2mm.

This is a very simple first draft of this issue; I will update the page with deeper explanation as soon as possible. Email me if you want to discuss this in any more detail.