Video observation enjoys increasing popularity among meteor observers. It has a number of advantages
over other techniques in the optical domain, which have been discussed in detail in .
One problem is the large amount of work required to reduce video observations. As in meteor photography, you need to measure the positions of meteor shutter breaks and a number of reference stars with high precision. In addition, finding meteors on video tapes is non trivial.
MetRec is a flexibel meteor detection software package developed by the author. It inspects a video data stream in real time and saves the appearance times of meteors. MetRec can also be employed to run a fully autonomous meteor observatory.
This paper discusses algorithms and features of MetRec. It places the software in the framework of other activities in this area and presents first results from evaluation tests.
First meteor observations with image intensified video systems were carried out by professional
astronomers in the late sixties  and by amateurs in 1986/87 
. In the eighties personal computers
became available. They supported the data reduction of video observations. However, due to restrictions
of processing power, computers were used for position measurements only. The automatic search for
meteors had been proposed early , but in practice all video tapes were inspected in painstaking manual
work until now.
First tests of computer software for real-time meteor detection were reported by the author in 1993 . Due to insufficient computing speed and problems with noisy first generation image intensifiers, the software had limited performance only.
The same algorithms were applied to recordings from second generation video cameras in 1996 . Even though several passes were still needed to inspect a video tape, the software achieved a good performance. The detection rate was about 75% on a 30 minute test recording with 28 meteors.
Parallel to the activities of the author, the first automatic meteor observatory (AMDES) was developed and tested by Pete Gural in the U.S. . There were two main differences between these systems:
The basic detection algorithm is similar to the one described in detail in . At the front end, half
resolution PAL (*) video frames (384x288 pixel) are digitized at a rate of 10-15 frames per second. The
frames are differentiated, i.e. the previous is subtracted from the current frame. This removes all stationary
and slow moving objects like stars from the image.
Next the difference image is flatfielded to obtain constant sensitivity in the full field of view. Contrary to the old algorithm, the image is divided by the flatfield (instead of subtraction). Division lookup tables are used to accelerate this computationally expensive operation. The flatfield itself is dynamically derived from the video stream.
To enhance longish objects like meteors in the noise difference image, pixel sums over 5 neighbours in 8 different directions are computed (see  for details). Contrary to the old approach, this is not done for all pixels, but for a few ten regions of interest (i.e. bright pixels) only. Again, this saves computing time and allows complex algorithms to be executed in real-time.
A meteor is detected, if one or more of these regions exceed a certain detection threshold. To gain high meteor detection rates without false alarms, the threshold is dynamically adapted to the current noise level.
Problematic with this kind of differential detection algorithm are rapidly changing objects like bright stars behind trees or a clock superimposed to the video image. MetRec handles these artifacts in a flexibel way. You can supply a b/w bitmap where all regions suitable for meteor detection are marked. The program reads this bitmap on startup and masks all other regions at run time. Masking does not help in case of flashing airplanes and fast moving satellites, however.
MetRec can be used for automatic observation without a VCR. You can configure the software such that it saves a fixed number of video frames and / or a sum image each time a meteor is detected. Saving is done at the full rate of 25 frames per second with only one or two frames lost at the beginning.
The performance of MetRec is mainly controlled by the detection threshold. This threshold is dynamically adapted to the current noise level to achieve best performance. However, you can control the working point with a recognition rate factor. A large factor will cause only few false detection, but you may also miss a number of faint meteors. Smaller values near one will raise the number of recognized meteors, but the number of false detections increases rapidly at the same time.
A major difference between the differential approach described here and the Hough transform based
algorithms is the rate of meteor detection. In the first case the algorithm is fast enough to decide for each
video frame whether or not a meteor is visible. The Hough transform, however, is computationally more
expensive. Here a number of video frames are stacked together before they are inspected using the
MetRec can be configured in a similar way. The software can merge a number of video frames before the sum image is differentiated and meteor detection starts. This increases the number of inspected video frames, as the time consuming detection is carried out after a number of video frames only. On the other hand the time resolution decreases. This results in less accurate values for meteor velocity and direction.
The computer's internal clock is heavily accessed during recognition. This may result in a considerable
clock skew on some machines. To avoid this problem, MetRec has the option to adjust the internal clock
with an external DCF-77 time signal receiver connected to the serial port.
Furthermore you can adjust brightness and contrast of the video signal and accelerate or slow down the
software to adapt to the underlying hardware. All options are set in a configuration file which is read at
startup time. The output is written to a log file.
An overview of the detection algorithm is given in Figure 1.
MetRec was designed for the frame grabber family of Matrox. It is written in the C programming language
using the Matrox Imaging Library MIL. MetRec requires a Pentium PC with at least 133 MHz clock rate
and 4 MB of main memory. The software runs under Dos, which is widely available and consumes much
less CPU time than other operating systems.
MetRec is shareware. Amateur astronomers may obtain it at no cost. Professional or commercial users need to register with the author first. If you intend to improve the code, to adapt it to your own hardware or to implement similar detection algorithms, you may also obtain the source code at request.
A number of preliminary performance tests have been carried out with MetRec. The software was running
on a 200 MHz Pentium MMX machine with a Matrox Meteor I PCI frame grabber (maximum data
transfer rate: 45 MB/s). The frame rate with this configuration was about 13 frames per second without
Evaluation tests were carried out on 1 hour of video recordings from the 1996 Geminid maximum. The observation was done with AVIS, a video system consisting of a 0.75/65 mm lens, a second generation image intensifier and a CCD video camera with manual gain control. The field of view was about 20° in diameter with a limiting magnitude for stars of about 8.5 mag. The observation was stored on VHS video tape.
The tape was difficult to analyse, as the lens was not slowed down during observation. All bright stars are burned out and show a strong coma if located near the edges. Meteors sometimes appear fuzzy and unfocussed, too. So the numbers reported here may be regarded as lower limits.
During a single pass visual inspection 44 meteors recognized. In addition there was 1 false detection, i.e. after replay of a suspicious event no meteor was found. After a number of test runs, an overall of 47 meteors were found. This gives a human detection performance of about 95% with 1 false alarm per hour on this test set.
MetRec was run with different recognition rate factors and frame stacking parameters. The results are shown in Figure 2 and 3. With a save recognition factor of 1.2, the detection rate varied between 60 and 75% (depending on the number of stacked frames) with false rates of 0 to 3 detections per hour. As was expected, the recognition rate increased to values between 75 and 85% when the recognition rate factor came closer to one. However, at the same time the number of false alarms increased dramatically to inacceptable values between 10 and 80 detections per hour.
3 faint meteors (5%) were never found in any of the tests.
Considering a population index in the order of 2.5 for the mixture of Geminids and sporadic meteors, the detection rate can be interpreted as a loss of limiting magnitude of roughly 0.5 mag.
MetRec is a software package for single-pass real-time detection of meteors. It results from
developments that started already in 1993. Preliminary tests under difficult conditions have
shown, that MetRec finds about 70% of all recorded meteors, which compares to human
performance of about 95%. This is similar to a decrease of limiting magnitude by half a
Contrary to AMDES, another software package for the same purpouse, a more simplistic algorithm is used. In the near future, Peter Gural and the author intend to evaluate both programs on the same recordings, which is necessary for a realiable comparison of the different algorithms .
Beeing shareware, MetRec is available to other video observers without restrictions.