Anchor-Free Target Detection via Probability Update

The patent extracts and removes the problematic preset anchor box mechanism from the detection system. Instead of using predetermined anchor boxes that generate many invalid candidates, the invention directly identifies potential target regions through detection points and probability values, eliminating the computational overhead of processing numerous invalid anchor boxes while maintaining detection coverage.

The patent changes the fundamental parameters of the detection approach by shifting from a fixed anchor box framework to a dynamic probability-based detection point system. This parameter change allows the system to adapt to different target densities and distributions without being constrained by predetermined box sizes and positions, thereby reducing computational waste while preserving detection effectiveness.

The patent introduces dynamics into the detection process by implementing an iterative probability update mechanism. Instead of statically removing high-probability points, the system dynamically adjusts probability values based on distance thresholds and iterative refinement, allowing it to adapt to dense target scenarios where targets may be closely spaced. This dynamic approach maintains processing speed while improving detection accuracy in challenging scenarios.

The patent incorporates feedback mechanisms through iterative probability updates. Detection points that initially appear to be false positives can have their probabilities re-evaluated in subsequent iterations based on their spatial relationships with other detected targets. This feedback loop ensures that valid targets in dense scenarios are not incorrectly discarded, thereby maintaining detection accuracy while preserving processing efficiency.