Current video surveillance systems transmit images and video sequences from the cameras to the screens in security control rooms. There the personnel can watch a limited number of selected screens. But, what happens in front of all the other cameras that are not linked to a screen at this moment? Automatic recognition of unusual or dangerous events and immediate triggering of an alarm could improve security and shorten reaction times. This would need smart systems using image processing and semantic networks.
Image processing software will find certain patterns or objects like persons or vehicles, identify them, track them through moving scenes, determine their size and location, or read inscriptions. By application of machine learning known patterns, images, objects or movements are “trained” for later recognition by the surveillance system. The semantic network understands the meanings of the actions in front of one or more cameras, interprets it and automatically derives decisions.
The challenges in research and development are
- further improvement of the reliability of image processing algorithms in consideration of extreme lighting conditions, partial covering, shade, reflections, etc.
- increasing efficiency of the image processing algorithms regarding computing time and storage requirements
- machine learning
- automatic reasoning and decision taking
- embeddeding the software in optimised co-operation with dedicated embedded signal processing hardware
heinrich.garn@ait.ac.at