Genetic Algorithm-Assisted Feature Extraction and Selection for Global Motion Estimation

The extraction and selection of good features for tracking is critical to the robustness of Global Motion Estimation, with application in several areas including video stabilization. The classical approach to this problem involves first extracting real-world points, based on structural criteria such as edges and corners, and subsequently selecting the more reliable features for tracking. Potential information in the movements of non-structural elements could thus be lost during feature extraction, while the selection criteria may not correlate well with camera movements. We propose a genetic algorithm-assisted approach, in which the feature extraction-selection process is directly coupled to the robustness of global motion estimates. This adaptive approach effectively learns the feature set whose movements, most closely correspond to global motion, thus ensuring robustness. This method was tested in application to video stabilization, and in comparison with peer approaches, was found to yield enhanced stabilization.