Status: Robotic Rock / Meteorite Classification

Meteorite Classifier Report from Liam Pedersen

Work here is progressing slowly, but steadily. There have been a number of technical mishaps that have delayed things. Nevertheless, they seem to be solved and I expect trials to begin within 3-4 days.

Pirrit Hills and Martin Hills Field Trip

We footsearched both these areas, finding one meteorite at Martin Hills. Inclement weather and the extreme cold made digital photagraphy of rocks and spectroscopy with the current equipment impossible. The digital camera would not operate, even though kept close to my body for warmth.

Instead, I concentrated on foot search and rock sample collection. Approximately 20 rock samples that appeared to cover all the rock types in the area (mostly granite) were collected and individually labelled. Color slide photographs were taken of some of these, along with photographs of the area and the boulder fields where we searched. These images should be useful to train the next classifier.

Spectra of the samples have been taken back in Patriot Hills base camp, and our currently being used with last seasons data set to retrain the classifier for the coming tests.

Hi-Resolution Pan/Tilt Mounted Color Camera

High-Resolution Camera Condensation Problem
Moisture in heated camera enclosure on pan tilt mounted color camera condenses and freezes on front glass pane:

Probable cause is snow entering through small holes in camera enclosure, melting due to heaters, and condensing and freezing on cold front plate. Holes are very small, largest approx 1 cm2. The solution implemented involves a good amount of duct tape and rice. Use duct-tape to seal all openings, remove all interior snow, and put in bags of rice to absorb water vapour in enclosure. (No other dessicant was available).

This has worked so far. It should be mentioned that convenient camera access is need in future. This enclosure was almost TOO robust.

Camera Scaling Calibration Experiment 11/10/98
Angular field of view (fov) of camera pixels various zooms had to be deduced experimentally by imaging a calibration target of known dimensions at various distances from the camera and zoom settings. This information is combined with the distance from the camera to the image subject to compute the scaling (mm/pixel) of the acquired image and thus compute the dimensions of the subject.

Complications occur because the camera zoom is measured by an encoder that outputs numbers whos relationship to the actual zoom value (in mm) is unknown. Furthermore, distance to object not known exactly, but is deduced from direction camera is pointing, assuming that the robot is on flat ground.

Missing Color Channels
The camera occasionally only acquires 2 of the 3 color channels (R,G,B):

Currently, the blue channel is flaky (see figure above), typically 20 times less bright than the other channels. Suspicions are that the Mutech frame grabber is at fault, since cables have been replaced, different cameras tried. Furthermore, the corrupted channel is not consistent. Sometimes the red or green channels are corrupted instead.

The rock classification code may be modified to deal with the loss of the blue channel, at the expense of accuracy. It cannot be changed to cope with the loss of the other channels. Intentions are to use Kodak handheld digital camera to get supplementary images of all rocks for post validation of classification code.


I have made some minor changes to the spectrometer system to cope with some problems with the attenuator that cropped up, possibly due to damage in transit. There is also a problem with an erratic connector, so the system is unstable in the off robot configuration (a different connector is used on the robot). I should like to compliment John Bates on producing good and well documented code for the spectrometer that made it much easier to diagnose and correct the problem.

To do

Prior to commencing trials need to:

Return to Latest Updates