A cost-effective background subtraction technique.
Pereira, Alex Lopes
DATA DE PUBLICAÇÃO
Background Subtraction is a very important task in image processing because its results are used in algorithms that recognize more complex object behaviors. This proposed research technique extracts movement evidences from difference: 1) between two consecutive frames; 2) between current frame and the fourth previous frame; and 3) between the current frame and a background model. These evidences are combined using the strategy of adding complementary values before applying thresholds. This strategy, combined with the application of the "iterate only once" requirement, leads to a Cost-effective Background Subtraction Technique. The main contribution of this work is the development of a novel pixel classification metric. Besides, it was extended by the following incremental improvements: 1st) The proposition of a half-connected filter as a fullfilment of the "iterate only once" requirement; 2nd) The extension of a simple and efficient shadow filter; and 3rd) The development of a quick way to evaluate accuracy of background subtraction techniques, based on a Genetic Algorithm (GA) and a Distributed Processing environment. When compared to recent research, the proposed technique results are better in performance and accuracy, this last one is due to an optmization process using a Genetic Algorithm. When performing tests on an Intel Dual Core Pentium 1.60GHz microprocessor with 1GB RAM, up to 376 Frames Per Second (FPS) of 160x120 color images were classified using this technique.