Dual AI Model Crowd Counting with Annotation Error Correction

The patent segments the crowd counting task from traditional full-body object recognition. Instead of detecting entire human bodies with bounding boxes, the method divides the task into detecting specific key body parts (head, torso, limbs) separately. This segmentation allows the system to count individuals even when full bodies are occluded, as long as key parts are visible, thereby improving accuracy while maintaining implementation simplicity.

The patent transitions from two-dimensional bounding box detection to a more sophisticated multi-dimensional approach by tracking multiple body parts across different spatial dimensions and temporal frames. This dimensional expansion allows the system to reconstruct complete person identities from partial observations, resolving the contradiction between simple implementation and accurate counting under occlusion.

The patent implements self-service annotation where the AI model automatically generates initial annotations for training data. The system uses semi-supervised learning where manually annotated data is used to train the model, which then generates annotations for additional data. This self-annotation capability reduces annotator fatigue and improves consistency, as the model applies the same annotation criteria uniformly across all data without human variability.

The patent performs preliminary automated annotation before final manual review. The AI model pre-processes and generates draft annotations that are then refined by annotators. This preliminary action reduces the time and effort required for manual annotation while maintaining high accuracy, as annotators only need to correct rather than create annotations from scratch.

The patent implements feedback mechanisms in the training process where the model's predictions are continuously evaluated against ground truth data. Annotation errors are detected through this feedback loop, and the training process incorporates error correction by adjusting weights and parameters based on prediction performance. This allows the system to maintain training efficiency while progressively improving prediction accuracy by learning from and correcting annotation errors.

The patent employs dynamic training where the importance of different training samples is adjusted based on their annotation quality and the model's current performance. Samples with suspected annotation errors are down-weighted or re-examined during training. This dynamic approach allows the system to efficiently train on large datasets containing some errors while maintaining high prediction accuracy by adaptively focusing on reliable training data.