On-line flash recognition algorithm was developed by Marcin Sokolowski,
       its parameters were optimized by Lech Wiktor Piotrowski

The purpose of the flash recognition algorithm is to detect optical flashes. The flash is defined as event when signal is detected on new image and no signal is detected on set of previous images. The algorithm must be fast, it has to analyze new image when next exposure is being taken. After dark subtraction, custom designed "Laplace" filter is applied, the example of unfiltered and filtered image is shown on images below :

Image before and after Laplace filter
            Image before filter                Image after filter


Stars on image after applying filter are much more clearly visible. The average and rms of ADU in pixels is determined and signal threshold Tn= 5 σ is calculated. The best values of parameters where tested by simulation of optical flashes and they were determined by Lech Piotrowski. The efficiency of flash identification was also tested and for flashes of brightness 8m-10m  it is ~80%. More information on simulation and algorithms parameters testing can be found here. The following subsequent criteria were applied to reduce the amount of background events and preserve as much real events as possible :


Several cosmetics cuts which clean the list of candidates from obvious background events are applied, these are :
Events which satisfy all the criteria are accepted as flash candidates from single camera. They are in most cases caused by cosmic rays events
hitting the CCD chip and generating flash-like events, example of cosmic-ray event is shown on image below :

example of cosmic ray event

The flash like event appears only on a single camera and is absent on the second camera observing the same field on the sky. In order to reject such kind of events, lists of events from both cameras are checked and only events visible on both cameras in the same position ( Distance < Rcoinc ) are accepted as real astrophysical events. In case only one camera is working, this condition is replaced by requirement that event is visible on at least 2 consecutive images. The coincidence cut rejects all background caused by cosmic ray events. The remaining events are mostly caused by flashing satellites, planes, meteors and sometimes constant stars. In order to reject those, the following cuts were implemented :

Examples of background events can be seen here. For events which survived all the above cuts sub renders of images are created and events are saved to the database. The local database is synchronized every 5 minutes and after this time events can be inspected by web browser. Most of them are still background events which were not rejected by the algorithm, but they can be easily rejected by human eye inspection. Events which could not be rejected by any criteria are considered as optical flashes caused by astrophysical processes and they are published on WWW :
The biggest problem of the flash recognition algorithm working on data from the prototype in LCO are flashing satellites. It is impossible to definitely reject events caused by this kind of objects due to the fact that orbital elements catalog is not complete and not all satellites flash often enough to reject them by track fit. In many cases satellite enters the FOV of the camera and flashes only once. In the final setup of the "Pi of the Sky" experiment, cameras observing same field of the sky will be placed in different locations separated by distance of ~50-100 km. Such configuration will allow to reject flashes caused by close satellites using the parallax. The parallax method was already tested when "Pi of the Sky" optical flashes were verified on RDOT experiment images, for more details see here.

mail Marcin Sokolowski 05.03.2007