Moving Object Detection Model for Afterimage-Blurred Images

The patent transitions from conventional two-dimensional image training to three-dimensional virtual space modeling. By constructing 3D models of moving objects and backgrounds, and projecting them onto 2D planes with controlled parallax effects, the system generates training images that accurately represent afterimage phenomena. This dimensional approach allows the detection model to learn the characteristic patterns of moving objects even when they appear as blurred afterimages in captured images.

The patent creates virtual copies of moving objects and backgrounds in a three-dimensional virtual space. These virtual models are then projected and synthesized to generate training images that replicate real-world afterimage conditions. By using synthetic copies rather than requiring extensive real captured images with afterimages, the system efficiently generates diverse training data without needing complex data collection processes.

The patent implements automated annotation through computer graphics technology. The system automatically specifies the positions of moving objects in generated training images by tracking the movement trajectories of virtual models in the three-dimensional space. This self-annotation process eliminates the need for manual labeling, significantly reducing data preparation time while maintaining high training data quality through precise automated position specification.

The patent performs preliminary positioning of virtual moving objects in the three-dimensional virtual space before generating training images. By pre-defining movement trajectories and positions of virtual models, the system automatically determines the correct annotation positions in the projected 2D training images. This preliminary action in 3D space streamlines the annotation process and eliminates time-consuming manual labeling.

The patent creates a universal three-dimensional virtual space framework that can model various types of moving objects and backgrounds. This unified 3D environment serves multiple functions: it generates diverse training images with different object positions, handles various movement patterns, and produces consistent annotations. The multi-functional virtual space approach enables comprehensive training data generation without requiring separate complex systems for each scenario.

The patent implements dynamic movement of virtual objects within the three-dimensional virtual space. By defining movement trajectories and allowing objects to move along these paths, the system automatically generates training images with diverse object positions. This dynamic approach efficiently creates varied training data while maintaining manageable system complexity through parameterized movement control.