Pixel Tracking via Anchor Frame Segmentation

The patent segments the trajectory formation process by introducing multiple anchor frames at different time points. Instead of concatenating optical flow continuously between consecutive frames, the method divides the trajectory into multiple segments, each anchored to a specific anchor frame. This segmentation reduces error accumulation within each segment while maintaining implementation simplicity.

The patent applies preliminary action by pre-identifying multiple anchor frames throughout the video sequence before trajectory formation. These anchor frames serve as predetermined reference points that stabilize the trajectory estimation process. By having these reference points established in advance, the method prevents drift accumulation without requiring complex real-time corrections.

The patent changes the parameter of temporal spacing between anchor frames. By strategically positioning anchor frames at varying intervals throughout the video sequence, the method adapts to different scene conditions. In complex scenes with occlusions, anchor frames are positioned to provide better coverage, while in simpler scenes, the spacing can be reduced to maintain low computational complexity.

The patent introduces anchor frames as intermediary reference points between the start and end of trajectories. These intermediary frames act as mediators that break the direct long-term tracking into shorter, more reliable segments. Each anchor frame serves as a local reference that is less susceptible to occlusion effects, thereby improving robustness without significantly increasing computational complexity.