Analyzing Surveillance Videos using automatically generated processing sequences with Knowledge-Augmented Genetic Algorithms

Extracting meaningful information from large number of video streams require designing specific algorithms to detect each type of object such as faces, people, vehicles, bags etc. The development of such specific algorithms requires a large amount of time from an expert in image analysis. Optimization based techniques have been increasingly used to automatically develop such algorithms, but they do not utilize any domain knowledge. Consequently, these automated approaches explore a large solution space and were only able to use a small number of primitive tools as building blocks in the generated algorithms. We proposes a novel method which integrates abstract knowledge about image processing tools into a genetic algorithm by exploiting the fact that there are classes of image processing algorithms that implement specific categories of algorithms such as noise reduction, sharpening, edge detection, binarization, classification etc. Using such knowledge, we were able to constrain the search performed by the genetic algorithm within a rich space of possibly successful processing sequences. Moreover, the use of abstract knowledge decouples the proposed method from implementation details of specific processing tools so that the system can be easily extended by incorporating additional tools. Experimental evaluations compare the our approach with a traditional genetic algorithm based implementation which does not utilize high-level knowledge. A case study shows that the proposed method could converge to the optimum solution six times faster than the traditional method.