LiDAR Localization Using RNN for Temporal Smoothness

The patent replaces traditional handcrafted localization pipelines with a neural network-based system. Specifically, it uses a U-Net architecture for cost volume processing and RNNs for temporal smoothing, substituting manual feature engineering and geometric algorithms with learned representations that automatically adapt to different scenarios without requiring substantial engineering tuning.

The patent transforms the localization problem into a parameter estimation task where the neural network learns optimal parameters (localization offsets) from training data. By changing from fixed handcrafted parameters to learned parameters, the system achieves high accuracy across diverse scenarios while reducing the need for scenario-specific engineering adjustments.

The patent replaces scenario-specific handcrafted pipelines with a universal neural network system. The U-Net and RNN architecture learns generalizable features from training data across multiple scenarios, enabling the same system to adapt to different environments (urban, rural, daytime, nighttime) without requiring scenario-specific tuning or multiple specialized pipelines.

The patent performs preliminary learning during a training phase where the neural network processes labeled data from various scenarios. This preliminary action allows the system to pre-learn scenario-invariant features and adaptations, so that during actual operation, the same trained model can generalize to new scenarios without requiring additional scenario-specific engineering work.

The patent merges semantics processing and 3D geometry localization into a unified neural network framework. The U-Net processes the cost volume (which encodes both semantic and geometric information) and the RNN processes temporal sequences, combining the strengths of neural networks in both semantic understanding and geometric parameter estimation within a single end-to-end trainable system.

The patent creates a composite neural network architecture that combines different network types (U-Net for spatial processing, RNN for temporal processing) to handle both semantic and geometric aspects of localization. This composite structure leverages the complementary strengths of each network type, with U-Net excelling at spatial feature extraction and RNN at temporal smoothness, achieving superior performance in both semantics and 3D geometry tasks.