Hierarchical Motion Estimation Search To Avoid Local Optima

The patent segments the full search process into multiple hierarchical stages: first performing a limited search in a small rectangular window around the initial motion vector, then progressively expanding the search range in subsequent stages only if improvement is detected. This segmentation transforms the single overwhelming full search into manageable sequential steps, reducing immediate computational complexity while preserving the ability to achieve optimal accuracy when needed.

The patent implements dynamic search range adjustment based on actual search results. The search window size and position are dynamically modified during the process - expanding when improvements are found and contracting when no further improvement is possible. This dynamic adaptation allows the system to balance between computational efficiency and accuracy, avoiding unnecessary searches in regions unlikely to contain better motion vectors.

The patent performs preliminary actions by first conducting a quick limited search in a small rectangular window around the initial motion vector prediction. This preliminary search establishes a baseline cost value and identifies a starting point for subsequent searches. By performing this preliminary action, the system prepares the search state in advance, enabling faster convergence in later stages while avoiding the need to search the entire range from scratch.

The patent implements feedback mechanisms where each search stage evaluates the cost function at candidate points and uses this feedback to determine the next search configuration. If a better match is found, the search range is expanded around the new best point; if no improvement is found, the search terminates early. This feedback-driven approach ensures that the system adapts its search behavior based on actual results, maintaining accuracy while improving speed compared to exhaustive full search.

The patent dynamically adjusts the search range based on the progression of the search process. The initial search is confined to a small rectangular window around the predicted motion vector. Only when improvements are detected does the system expand the search range in subsequent stages. This dynamic range adjustment ensures comprehensive coverage when necessary while minimizing search time in cases where the predicted motion vector is already accurate.

The patent segments the search range into multiple hierarchical levels: a small initial rectangular window, followed by progressively larger search areas in subsequent stages. This segmentation allows the system to focus computational resources on the most promising regions first, achieving effective coverage with minimal search time. The segmented approach ensures that even if the initial prediction is inaccurate, the system can systematically expand coverage without searching the entire range from the beginning.