Video Behavior Recognition Model Using Selective Feature Combinations

The patent changes the fundamental parameters of behavior recognition from rule-based joint position patterns to machine learning models that process diverse features including joint positions, velocities, accelerations, and temporal patterns. This allows the system to adapt to complex behaviors by learning from data rather than relying on predefined rules

The system transitions from static rule-based patterns to dynamic machine learning models that can adapt and learn from new behavior patterns. The models are trained on diverse datasets and can continuously improve recognition accuracy for complex behaviors through exposure to more training data

The patent performs preliminary actions by pre-training models on comprehensive datasets that capture diverse behavior patterns. Once trained, the models can recognize complex behaviors with high accuracy without requiring extensive additional sample preparation for each specific behavior recognition task

The system uses synthetic or augmented training data that copies and transforms existing behavior patterns to create diverse training samples. This reduces the need for extensive manual sample collection while maintaining recognition accuracy through data augmentation techniques