A Novel Motion Estimation Technique Using Genetic Algorithm Search

Traditional motion estimation (ME) techniques have relied upon the assumption that their evaluation function was sufficiently unimodal to warrant the application of simple gradient based search to find the displacement vector field or their image sequence. What we have done is to iDVF) evelop a more robust M E technique which uses Genetic Algorithms (GA) to maintain a statistically generated p o p ulation of candidate solutions’. Our ME technique works within a complex motion environment containing three dimensions of displacement and rotation and thus requires 6 degrees of freedom. These 6 degrees of freedom are implemented using a novel frequency domain image warping technique which reestablished a frame to frame correspondence and allows for the application of a correlation measure as the fitness function. T h e paper presents a discussion of the advantages of GAS for multimodal search in the context of motion estimation in a complex environment and presents a novel means of hybridizing the search so as to improve the convergence properties of the algorithm. Simulation results are used to show the performance of GAS in locating global solutions to the ME problem.