Radar Point Cloud Generation Without Precise Sensor Synchronization

The patent replaces the conventional mechanical/synchronization-based approach (requiring common oscillators and optical signals) with a machine learning model (transformer-based neural network) that automatically processes radar data from distributed sensors. The model learns to merge data from sensors with varying positions and timing without requiring precise synchronization hardware, thus substituting complex mechanical synchronization systems with an intelligent software-based solution.

The patent implements a dynamic solution where the transformer-based model continuously adapts to changing sensor positions during platform operation. Unlike static conventional methods that assume fixed sensor positions, the neural network dynamically processes and merges radar data even when sensors shift due to vibrations, making the system robust to positional variations without requiring active compensation mechanisms.

The patent fundamentally changes the approach from maintaining fixed positional parameters to using a neural network that learns optimal data merging strategies across varying positional parameters. The transformer model is trained to handle a range of sensor positions and configurations, allowing the system to maintain high accuracy regardless of actual sensor locations, thus making the system adaptable to positional variations without requiring precise constant position knowledge.