Ultrasonic Classifier Tuning for False-Positive Obstacle Detection

The patent replaces rule-based empirical classifiers with machine learning-based classifiers. Specifically, it uses trained neural networks and support vector machines that automatically learn classification boundaries from training data, substituting the manual rule-based mechanical approach with an adaptive learning-based system. This improves classification accuracy while maintaining computational efficiency through optimized model architectures.

The patent transforms the classification approach by changing from fixed rule-based parameters to learned parameters from training data. The system learns optimal decision boundaries, feature weights, and classification thresholds from labeled training examples, allowing the classifier to adapt to different object types and conditions without manual rule adjustment.

The patent enables adaptability to different vehicle variants by training separate machine learning models for each vehicle type or by implementing a hierarchical classification system that adapts parameters based on vehicle configuration. The system learns vehicle-specific characteristics from training data, allowing easy adaptation to new vehicle variants through retraining with appropriate datasets without redesigning the entire classification system.

The patent performs preliminary training of classification models with vehicle-specific data before deployment. By pre-training models with labeled data from specific vehicle variants, the system captures vehicle-specific acoustic characteristics and environmental conditions, enabling accurate classification when the vehicle is actually in use without requiring real-time adaptation.

The patent segments the classification problem into hierarchical levels or modular components. Instead of requiring a single classifier to handle all object classes simultaneously, the system divides classification into stages (e.g., first distinguishing between stationary and mobile objects, then further classifying each category) or uses multiple specialized classifiers for different object types, reducing the complexity dependence on the total number of classes.

The patent develops universal machine learning classifiers that can handle multiple object classes through a unified framework. The trained models learn generalizable features and patterns that apply across different object types, allowing a single classifier architecture to perform multiple classification functions without requiring separate specialized systems for each object class.

The patent utilizes the vibrational characteristics of ultrasonic echoes to distinguish between different object types. By analyzing the frequency content, amplitude modulation, and temporal patterns of reflected ultrasonic signals, the system extracts features that differentiate stationary vehicles from other obstacles. Machine learning classifiers then process these vibrational features to achieve accurate classification.

The patent introduces machine learning classifiers as an intermediary between raw ultrasonic sensor data and final obstacle identification. The classifier acts as a mediator that processes the acoustic signals, extracts meaningful features, and translates them into accurate object classifications, preventing information loss that would occur with simple threshold-based detection.