Intersection Contention Detection Using Unified Neural Segmentation

The patent combines multiple separate deep neural networks that detect different features (traffic lights, stop signs, vehicle positions, lanes, free-space boundaries) into a single unified neural network. This unified network processes sensor data and directly outputs intersection areas and classifications, eliminating the need to separately detect and combine multiple features. The merging reduces device complexity and computational overhead while maintaining detection accuracy.

The patent employs data augmentation techniques to generate synthetic training data by copying and transforming existing annotated intersection data. This creates additional training examples without requiring manual annotation of new geographic regions. The copying approach enables scalable training data preparation while maintaining classification accuracy, as the synthetic data preserves the essential characteristics of real intersections.

The patent enables the vehicle's sensor system to independently detect and classify intersections in real-time without relying on pre-stored high-definition maps. The neural network processes current sensor data (camera, LIDAR, RADAR) to identify intersection areas and determine vehicle traversal priority directly from the environment. This self-service approach provides adaptability to transient conditions such as police-directed traffic or stopped school buses that are not reflected in static maps, while maintaining reliability through real-time environmental perception.