Radar-Camera Fusion for Spurious Object Detection in AVs

The patent combines radar data and camera images into a fused representation that integrates strengths of both sensors. The radar provides speed and range information while the camera provides visual confirmation, together resolving ambiguities that either sensor alone cannot distinguish, thereby improving detection accuracy without requiring completely separate independent systems.

The patent introduces machine learning models as intermediary components that process and interpret the fused radar-camera data. These models act as mediators between raw sensor inputs and final detection decisions, learning to distinguish genuine objects from spurious returns through trained patterns, thereby improving accuracy while managing system complexity through standardized processing layers.

The patent applies partial action by using machine learning models to selectively process only the most ambiguous or uncertain radar returns that require further verification. Rather than thoroughly analyzing every single return point, the system focuses computational resources on cases where distinction between genuine and spurious returns is most critical, maintaining speed while improving accuracy for problematic detections.

The patent replaces traditional mechanical or algorithmic methods of filtering radar returns with machine learning-based classification. Instead of using fixed thresholds or geometric filtering rules, the system employs trained neural networks that can rapidly evaluate complex patterns in the fused radar-camera data, achieving both high accuracy and fast processing speeds through learned decision boundaries.

The patent segments the classification task into distinct stages: initial radar detection, camera image processing, feature extraction from both modalities, fusion of features, and final classification. This segmentation allows each component to be optimized independently and enables parallel processing of different data streams, reducing overall computational complexity while maintaining high classification accuracy through specialized processing at each stage.

The patent performs preliminary processing of radar and camera data before fusion, including feature extraction, filtering, and preprocessing operations. By preparing and organizing data in advance, the system reduces the computational burden on the final classification stage, allowing machine learning models to work with pre-processed, structured inputs that require less intensive computation while still achieving high accuracy.