Neural Network Scoring for Autonomous Vehicle Trajectory Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training scoring models for autonomous vehicles to accurately assign confidence scores to predicted future trajectories of agents is challenging due to insufficient training signals, especially when the model has an excessive number of parameters, leading to poor prediction accuracy and training efficiency.
Innovation Solution
Implementing a neural network system that jointly trains multiple sub-networks to perform tasks such as generating intermediate representations and predicting future trajectories, providing richer training signals to improve the scoring model's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the scoring model has an excessive number of parameters to handle complex trajectory predictions, then the model's capacity to represent complex behaviors increases, but training signals become insufficient and training efficiency deteriorates
Solution Approach 1:
The patent segments the scoring model into multiple independent sub-networks, each responsible for predicting specific trajectory outcomes. This segmentation allows each sub-network to have fewer parameters and receive more focused training signals, while collectively maintaining the overall model's capacity to handle complex trajectory predictions.
Solution Approach 2:
The patent creates multiple sub-networks that share the same architecture and can be applied to different trajectory predictions. Each sub-network serves a specific function (predicting a particular trajectory outcome) while following a universal design pattern, allowing the system to maintain versatility through composition rather than through a single large model.
2Adaptability or versatility
If the scoring model has an excessive number of parameters, then the model can capture more complex patterns, but prediction accuracy deteriorates due to insufficient training signals
Solution Approach 1:
By dividing the model into sub-networks, each sub-network focuses on predicting specific trajectory outcomes with dedicated training signals. This segmentation ensures that each component receives sufficient training data and signals to learn accurate patterns, while the aggregation of multiple sub-network predictions maintains overall pattern capture capability.
Solution Approach 2:
The patent implements a training mechanism where ground truth trajectory data provides feedback signals to update the sub-network parameters. This feedback loop ensures that each sub-network learns from actual outcomes, improving prediction accuracy through iterative refinement based on measured performance gaps between predictions and ground truth.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a neural network having a plurality of sub neural networks to assign respective confidence scores to one or more candidate future trajectories for an agent. Each confidence score indicates a predicted likelihood that the agent will move along the corresponding candidate future trajectory in the future. In one aspect, a method includes using the first sub neural network to generate a training intermediate representation; using the second sub neural network to generate respective training confidence scores; using a trajectory generation neural network to generate a training trajectory generation output; computing a first loss and a second loss; and determining an update to the current values of the parameters of the first and second sub neural networks.