Optical Flow Fusion for Large Motion and Small Object Accuracy

The patent segments the image into patches and processes them in a hierarchical manner through multiple levels of an image pyramid. Instead of exhaustively searching all patches at full resolution, the algorithm divides the problem into coarse-to-fine stages, where each level processes smaller regions first and refines results at progressively finer resolutions, significantly reducing computational complexity while maintaining matching accuracy.

The patent performs preliminary actions by first computing optical flow at coarse levels of the image pyramid to establish initial motion estimates. These preliminary results are then used to guide and constrain the exhaustive search at finer levels, so that the computationally expensive precise matching is only performed in relevant regions with reduced search spaces, rather than exhaustively searching the entire image at full resolution from the start.

The patent employs a dynamic, multi-scale approach using an image pyramid where the analysis resolution adapts to the magnitude of motion. At coarser levels, the algorithm can capture large displacements that would be impossible to detect at fine resolutions. The spatial coherency constraint is dynamically applied at each level, allowing the system to maintain stability for small motions while adapting to handle large motions through the hierarchical structure.

The patent adds a dimensional aspect by introducing multiple resolution levels through the image pyramid. Instead of working in a single fixed resolution space, the algorithm operates across multiple scales, transforming the problem from a two-dimensional spatial matching task into a multi-dimensional problem that includes resolution level as an additional dimension. This allows large motions to be captured at coarser levels while maintaining precision at finer levels.

The patent segments the image processing task across multiple resolution levels, with each level handling different scales of features. Coarse levels capture large-scale motions and structures, while fine levels preserve small details and thin objects. By segmenting the problem this way, the algorithm ensures that small or thin objects are not lost in the down-sampling process, as they are processed and matched at finer resolution levels where their details are preserved.

The patent introduces the resolution level as an additional dimension, creating a multi-scale processing framework. Instead of losing information through single-level down-sampling, the algorithm maintains information across multiple dimensional levels. Small or thin objects that may be obscured at coarse levels are recovered and properly matched at finer levels, effectively using the resolution dimension to preserve detail information that would otherwise be lost.

The patent performs preliminary action by first computing a coarse-level optical flow field that captures large-scale motions. This preliminary flow estimate is then used to initialize the fine-level optimization, replacing the conventional zero initialization. This preliminary computation at coarse levels provides a head start for handling large motions, while the overall algorithm structure remains relatively simple and builds upon conventional optical flow methods.

The patent makes the initialization dynamic rather than static. Instead of always initializing to zero, the algorithm dynamically computes an appropriate initialization based on coarse-level analysis. This dynamic initialization adapts to the actual motion magnitude in the image pair, allowing the simple conventional optical flow framework to handle both small and large motions effectively by adjusting its starting point based on the problem characteristics.