Siamese Neural Network Data Association for Digital Map Updates

The patent replaces traditional mechanical data association methods (feature matching, point cloud registration) with a machine learning-based neural network system. The neural network learns optimal association patterns from training data and automatically performs data association, achieving both high precision in position indication and high processing speed for frequent map updates.

The patent changes the approach from deterministic parameter-based association to probabilistic learning-based association. By training the neural network on labeled data containing ground truth positions, the system learns to associate points with objects based on learned patterns rather than simple geometric or intensity thresholds, improving both accuracy and speed.

The patent implements a self-training mechanism where the system uses its own predictions combined with ground truth data from probe vehicles to continuously improve its performance. The neural network is trained on collected data and then deployed to make predictions, creating a self-improving system that achieves high reliability without requiring increasingly complex infrastructure.

The patent performs preliminary training of the neural network offline using collected training data before deployment. This preliminary action allows the system to learn optimal association patterns in advance, enabling it to achieve high navigation reliability during actual operation without requiring complex real-time processing infrastructure.