Large Language Model Prompts for Congested Pedestrian Trajectory Prediction

The patent replaces traditional mechanical trajectory prediction systems with a large-scale language model-based system. The LLM processes natural language descriptions of pedestrian contexts and historical trajectory data to predict future positions, substituting conventional algorithmic approaches with a neural network model that achieves superior accuracy in complex environments.

The patent introduces natural language descriptions as an intermediary between raw sensor data and trajectory prediction. The system converts environmental observations and pedestrian contexts into language-based prompts that the LLM processes, enabling more flexible and accurate reasoning about pedestrian intentions compared to direct coordinate-based methods.

The patent applies preliminary action by pre-training the large-scale language model on extensive language data before fine-tuning it for trajectory prediction. This pre-training establishes foundational language understanding capabilities that transfer to the specialized domain of pedestrian trajectory prediction, reducing the complexity of domain-specific training.

The patent applies local quality by fine-tuning the pre-trained LLM specifically for trajectory prediction tasks while maintaining the general language understanding capabilities. The model adapts its parameters locally for the prediction task without losing the beneficial general-purpose language processing abilities acquired during pre-training.