This paper describes a new, very efficient video object motion estimation strategy, that considers video object translation, translation of rotation center, and planar multilayering. It is based on a Multipopulation Modified Coevolutionary Genetic Algorithm (MMCGA), that receives the video objects of a segmented sequence of video images, and outputs the corresponding motion and layer information using appropriately represented object and layer genotypes. The algorithm performs a random search for locating the global optimal solution in the searching space. This operation involves an evolutionary process wherein populations of predicted solutions evolve over a period of generations. From the possible solutions, a video test frame is created, and the fitness of the test frames is evaluated by comparing with the reference frame. Next, solutions are selected for reproduction based on their fitness values. Good solutions are selected for reproduction while bad ones are eliminated. The selected solutions undergo recombination under the genetic operations of modified crossover and dynamic mutation. This last operation increases the mutation rate and reduces the mutation range from one population generation to the next, maximizing the performance of the MMCGA. For the increase in predicted solution accuracy, and convergence speed, lifetime strategies are used. Preliminary simulations with synthetic video images have shown very encouraging results with the proposed video motion estimation technique, which competes favorably with respect to the conventional motion estimation algorithms in accuracy, robustness, simplicity and speed.
展开▼
