Bounding Box Aspect Ratios for Occluded Object Coordinate Correction

The patent changes the parameter of aspect ratio threshold dynamically based on object type and occlusion probability. Instead of using a fixed threshold, the system adjusts the reference aspect ratio according to the detected object's category (human, animal, vehicle) and estimated occlusion level, allowing adaptive detection that maintains accuracy for occluded objects while preserving deep learning's overall detection capability

The patent replaces traditional geometric verification methods with a deep learning based occlusion detection mechanism. The processor uses neural networks to estimate occlusion probability and adjust aspect ratio thresholds, substituting mechanical/geometric rules with intelligent algorithms that can handle occluded objects more effectively

The system performs self-service by using its own detection results to guide further processing. The deep learning detector's output (bounding boxes and confidence scores) is immediately used to identify potential occlusions and trigger aspect ratio verification, creating a self-contained detection pipeline that doesn't require external databases

The patent introduces aspect ratio verification as an intermediary step between deep learning detection and final object confirmation. This intermediate verification layer filters out false positives from occluded objects without requiring additional training data or complex database queries, simplifying the overall system