PERSONALIZED VEHICLE LANE CHANGE MANEUVER PREDICTION

Smart overview of the Invention

The patent application discusses a system and method for predicting vehicle lane change maneuvers using a learning-based algorithm. This algorithm evaluates a driver's behavior to predict lane changes, utilizing both an online and offline phase. The offline phase involves training a machine learning model with historical vehicle data, while the online phase validates this model with real-time driving data to predict lane changes and potential vehicle trajectories.

Technical Context

This technology is relevant to the field of autonomous and connected vehicles, specifically in predicting lane change maneuvers. As autonomous vehicle technology advances, it becomes crucial for these systems to predict human-driven vehicle behaviors accurately to ensure safety and efficiency in mixed-traffic environments. The algorithm addresses the challenge of human driver unpredictability by personalizing predictions based on individual driving behaviors.

System and Methodology

The system uses a hierarchical algorithm that integrates online decision prediction with trajectory prediction, considering driver preferences and interactions. The offline phase trains a Long-Short Term Memory (LSTM) network with historical data, while the online phase uses this trained model to predict lane changes. The system refines its predictions by feeding real-time driver behavior data back into the training process.

Implementation

A vehicle control system is equipped with a personalized lane change prediction control circuit, consisting of memory and processors to execute machine-executable instructions. Key operations include obtaining historical driving data, training a machine learning model, predicting maneuvers using real-time data, and determining the most probable vehicle trajectory based on personalized cost functions. The system differentiates between lane change and lane keep maneuvers, applying clustering algorithms and morphological operations to enhance prediction accuracy.

Example Embodiments

Various embodiments of the system are described, such as generating training data through clustering algorithms and recovering personalized cost functions for lane change and lane keep maneuvers. The system determines feature weights for predictive features and selects the most probable trajectory based on these weights. Additionally, the model's predictive capacity is refined by continuously generating and incorporating new training data from real-time vehicle states.