Autonomous Vehicle Merge Scenario Detection and Yield Prediction

The patent segments the trajectory prediction problem into multiple discrete hypotheses (e.g., yield hypothesis, no-yield hypothesis, aggressive hypothesis, conservative hypothesis) for each interacting object. Each hypothesis represents a possible future behavior with associated likelihood. This segmentation allows the system to evaluate multiple potential trajectories simultaneously without requiring a single complex predictive model, thereby improving accuracy while managing system complexity through modular hypothesis evaluation.

The system dynamically adjusts the set of hypotheses and their likelihoods based on real-time observations of object behavior, contextual factors, and interactions. As new sensor data arrives, the system updates the probability distributions across different hypothesis scenarios, allowing the most likely trajectory to emerge dynamically. This dynamic adaptation improves prediction accuracy without requiring a statically complex system architecture.

The system generates multiple hypotheses for each interacting object (which is more than the single-trajectory approach), but applies partial evaluation by focusing computational resources on the most promising hypotheses based on initial likelihood assessments. Not all hypothesis combinations are fully evaluated to exhaustion; instead, the system prunes less likely scenarios early in the computation process, maintaining high prediction accuracy while reducing overall computational burden and processing time.

The hypothesis evaluation system uses its own output (likelihood scores from previous evaluation cycles) to guide subsequent computational efforts. High-likelihood hypotheses automatically receive more computational attention in the next iteration, while low-likelihood ones are deprioritized. This self-reinforcing mechanism allows the system to efficiently allocate computational resources without external intervention, improving processing speed while maintaining accuracy through iterative refinement.

The patent segments the complex multi-object interaction problem into discrete object-pairs, where each pair is evaluated independently for yield/no-yield hypotheses. This segmentation allows the system to manage safety considerations for multiple objects simultaneously by breaking down the overall complexity into manageable pairwise interactions, improving navigational safety while controlling system complexity through modular analysis of each object-pair relationship.

The system performs preliminary classification of object-pairs into yield and no-yield categories based on initial assessment of right-of-way rules, object positions, and motion patterns. This preliminary action establishes a structured framework for safety evaluation before detailed trajectory prediction occurs, allowing the system to prepare safety hypotheses in advance and reducing the complexity of real-time decision-making while maintaining high navigational safety standards.