Agent Trajectory Prediction Using Lane-Graph Path Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Predicting future trajectories of surrounding vehicles and pedestrians is challenging for autonomous vehicles due to inherent uncertainty, which affects safe and efficient navigation through complex traffic scenes.
Innovation Solution
A graph-based approach is used to encode high-definition maps and agent trajectories, conditioning predictions on lane-graph traversals, and employing sampled latent variables to account for longitudinal variability, resulting in diverse and precise trajectory predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional trajectory prediction methods are used, then computational simplicity is maintained, but prediction accuracy and reliability deteriorate due to inherent uncertainty in predicting future positions of agents
Solution Approach 1:
The prediction system segments the trajectory prediction task into multiple discrete trajectory hypotheses, each representing a plausible future path. Instead of predicting a single trajectory, the system generates and evaluates multiple segmented trajectory options, improving reliability by considering uncertainty through diversification of prediction outcomes
Solution Approach 2:
The system transitions from predicting position in 2D space to predicting trajectories in 4D space (x, y, time, probability). By adding the probability dimension and time dimension, the system captures uncertainty and temporal evolution of trajectories, significantly improving prediction reliability while managing complexity through structured probabilistic modeling
2Reliability
If comprehensive scene context is analyzed for accurate trajectory prediction, then prediction reliability improves, but computational time and processing resources increase
Solution Approach 1:
The system performs preliminary encoding of map features and agent features into compact representations before trajectory prediction. By pre-processing and encoding scene context into condensed feature vectors, the system reduces computational burden during the actual prediction phase while maintaining the reliability benefits of comprehensive scene analysis
Solution Approach 2:
The system replaces traditional mechanical search methods with neural network-based probabilistic modeling. Instead of exhaustively searching through all possible trajectories, the system uses learned probability distributions to directly generate likely trajectories, significantly reducing computational time while maintaining high reliability through the probabilistic framework
3Reliability
If multiple trajectory hypotheses are generated to account for uncertainty, then prediction reliability improves, but device complexity and computational resources increase
Solution Approach 1:
The system generates a limited number of discrete trajectory hypotheses (e.g., 3-5 most likely trajectories) rather than exhaustively sampling all possible trajectories. This partial action approach maintains reliability by focusing on the most probable outcomes while improving productivity by avoiding unnecessary computation of low-probability trajectories
Solution Approach 2:
The system uses parameterized trajectory representations with learnable parameters that capture essential trajectory characteristics. By changing from fixed-grid trajectory sampling to continuous parameterized representations, the system improves computing efficiency while maintaining reliability through learned probability distributions over trajectory parameters
Data Source
AI summary
Provided are methods for predicting agent trajectories, which can include generating a graph corresponding to a map of a scene by encoding map features and agent features as node encodings of the graph and determining a policy for application to outgoing edges of the nodes of the graph. Some methods described also include sampling paths for a target vehicle in the scene according to the policy and predicting a set of trajectories based on the sampled paths traversed by the policy and a sampled latent variable. Systems and computer program products are also provided.