Traffic Graph Behavior Prediction for Road User Intent Inference

The patent introduces a neural network model as an intermediary that processes sensor data, graph representations, and hidden variables to infer road user intentions and plans. This mediator translates observable data into predictive insights about future behavior, resolving the contradiction by enabling rich information exchange without requiring direct complex communication between road users.

The patent replaces the mechanical/physical turn signal system with a computational system using neural networks that process sensor data and graph representations. This substitution enables sophisticated behavior prediction without requiring physical changes to existing road user vehicles, achieving enhanced information exchange through software-based inference rather than hardware-based communication.

The patent segments the complex data processing task into distinct components: sensor data acquisition, graph representation construction, state variable extraction, and neural network processing. By dividing the data flow into manageable segments with specific functions, the system achieves reliable traffic situation assessment while keeping each processing stage manageable and modular.

The patent introduces graph representations as an intermediary data structure that organizes sensor data from multiple sources into a unified format with nodes representing road users and edges representing interactions. This intermediary layer simplifies the processing of heterogeneous data by providing a standardized structure that the neural network can efficiently analyze, reducing the apparent complexity of handling multiple data sources.

The patent replaces direct observation of hidden variables with computational inference using neural networks. Instead of requiring direct access to road user intentions and plans, the system uses the neural network to infer these hidden variables from observable sensor data and graph representations, achieving accurate behavior prediction through computational modeling rather than direct measurement.

The patent transforms the prediction problem by changing parameters from directly observable quantities to inferred hidden variables. The neural network model learns to map observable data (positions, velocities, graph structures) to hidden variables (intentions, plans, future behavior), effectively changing the parameter space to include unobservable but predictive quantities that improve prediction accuracy.