Synthetic HD Map Data Generation for Autonomous Vehicle Change Detection

The patent creates synthetic probe data by copying and transforming existing HD map data through simulated environmental changes. Instead of collecting real probe data from probe vehicles, the system generates artificial training datasets by programmatically modifying existing map features (adding/removing road markings, changing lane configurations) to represent various environmental change scenarios. This copying approach eliminates the time-consuming real-world data collection process while maintaining the precision needed for training change detection models.

The system performs preliminary actions by pre-generating synthetic training data and pre-training the machine learning model offline before actual HD map updates are needed. The synthetic dataset is created in advance with various simulated environmental changes, and the model is trained beforehand on this prepared data. When real updates are needed, the pre-trained model can quickly process new map data without requiring time-consuming real-time analysis or retraining.

The patent replaces complex physical probe data collection infrastructure with a digital copying system that generates synthetic data from existing HD maps. Instead of deploying probe vehicles with specialized sensors and processing equipment, the system uses software to copy and transform existing map data into synthetic training examples. This dramatically reduces device complexity while maintaining the ability to generate high-precision training data for change detection.

The patent substitutes mechanical probe data collection systems (probe vehicles, specialized sensors, physical data gathering infrastructure) with a computational system that generates synthetic data through algorithms. The mechanical process of driving probe vehicles around environments to collect data is replaced by digital transformation of existing map data, including programmatically adding/removing road markings, changing lane configurations, and simulating environmental variations through software operations.

The system performs preliminary training using synthetic data before actual HD map updates. The machine learning model is pre-trained offline on extensively generated synthetic datasets that cover various environmental change scenarios. This preliminary training phase prepares the model in advance, so when frequent updates are needed, the already-trained model can quickly process new map data and detect changes without requiring time-consuming retraining or analysis from scratch.

The system uses synthetic copying of environmental change scenarios to create comprehensive training data without requiring repeated real-world probe data collection. By copying and transforming existing HD map features into various simulated change scenarios (different road markings, lane configurations, environmental conditions), the system generates abundant training examples that enable the model to handle frequent updates efficiently without proportionally increasing real-world data collection time.