Trajectory Labeling With Intention and Outcome for AV Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for training autonomous vehicle trajectory prediction models rely solely on long-term labels that indicate actual outcomes, failing to capture the vehicle's intended actions, which leads to agnostic models that do not account for changing intentions due to environmental changes.
Innovation Solution
Generating training data that incorporates both short-term intention labels and long-term result labels to determine a ground-truth probability for candidate trajectories, allowing the model to consider what the agent intended to do and what actually happened, thereby improving prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If only long-term labels indicating actual outcomes are used for training, then the training process is simple, but the model becomes agnostic and fails to capture changing intentions
Solution Approach 1:
The training label is segmented into two distinct components: long-term labels indicating actual outcomes and short-term labels indicating initial intentions. This segmentation allows the model to learn both what agents ultimately did and what they initially intended to do, resolving the contradiction by making the training data more informative without overwhelming complexity
Solution Approach 2:
The approach adds a temporal dimension to the training labels by distinguishing between short-term intentions and long-term outcomes. This dimensional expansion transforms the training data from a single outcome-based label to a multi-temporal label structure, enabling the model to capture intention dynamics while maintaining manageable complexity through structured processing
2Measurement precision
If both short-term intention and long-term result labels are incorporated, then prediction accuracy improves, but the training data generation becomes more complex
Solution Approach 1:
The system performs preliminary classification of trajectory segments into intentional (short-term) and actual outcome (long-term) components before final model training. This preliminary action simplifies the overall process by pre-organizing the complex multi-temporal data into structured categories, enabling precise prediction while managing generation complexity through advance processing
Solution Approach 2:
The training data generation system acts as an intermediary that automatically processes raw trajectory data to extract both short-term intention labels and long-term outcome labels. This intermediary layer handles the complexity of multi-temporal label generation, shielding the model training process from complexity while delivering high-precision multi-dimensional training signals
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating training data for training a machine learning model to perform trajectory prediction. One of the methods includes: obtaining a training input, the training input including (i) data characterizing an agent in an environment as of a first time and (ii) data characterizing a candidate trajectory of the agent in the environment for a first time period that is after the first time. A long-term label for the candidate trajectory that indicates whether the agent actually followed the candidate trajectory for the first time period is determined. A short-term label for the candidate trajectory that indicates whether the agent intended to follow the candidate trajectory is determined. A ground-truth probability for the candidate trajectory is determined. The training input is associated with the ground-truth probability for the candidate trajectory in the training data.