Road Agent Trajectory Prediction Using Intermediate Space Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current prediction systems for vehicle trajectories, especially in autonomous and semi-autonomous vehicles, face challenges in accurately predicting the trajectory of road agents without requiring continuous retraining of neural networks when classification sets change.
Innovation Solution
A method involving a prediction system that calculates distances between intermediate representations of vehicle trajectories using a neural network, trains the network based on similarity indicators, and creates an intermediate space for predicting road agent trajectories, allowing predictions without needing to retrain the network even if classification sets are modified.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the neural network is retrained when classification sets change, then prediction accuracy may be maintained, but system efficiency and adaptability deteriorate due to continuous retraining requirements
Solution Approach 1:
The patent segments the feature space into multiple intermediate spaces, each corresponding to a specific classification set. The neural network learns to map inputs to these segmented spaces independently, allowing the system to adapt to different classification sets by selecting or switching between pre-trained intermediate spaces rather than retraining the entire network. This segmentation enables efficient adaptation while maintaining prediction accuracy.
Solution Approach 2:
The patent changes the parameter representation by introducing intermediate representations that are invariant to classification set changes. By transforming the original feature space into intermediate spaces with specific geometric properties (such as equidistant relationships), the system achieves parameter independence from classification labels, allowing accurate predictions without retraining when classification sets change.
2Productivity
If the neural network structure is simplified for faster processing, then system efficiency improves, but prediction accuracy deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network to create multiple intermediate spaces corresponding to different classification sets before actual prediction tasks. This offline preparation allows the network to be structurally optimized for speed while maintaining accuracy, as the complex adaptation work is done in advance rather than during real-time processing.
Solution Approach 2:
The patent introduces intermediate representations as mediators between the input data and final predictions. These intermediate representations simplify the processing by capturing essential features in a reduced space, enabling faster computation while preserving the information needed for accurate trajectory prediction.
3Adaptability or versatility
If multiple classification sets are supported, then system adaptability improves, but device complexity increases due to multiple neural networks
Solution Approach 1:
The patent implements universality by designing a single neural network that can serve multiple classification sets through the intermediate space mechanism. The network learns to map inputs to classification-invariant intermediate representations, allowing it to accurately predict trajectories across different classification sets without requiring separate networks for each set. This multi-functional approach reduces system complexity while maintaining high adaptability.
Solution Approach 2:
The patent adds another dimension to the feature space by introducing intermediate representations that incorporate classification set information as an additional dimension. This dimensional extension allows the single network to handle multiple classification sets by learning relationships in this expanded space, avoiding the need for multiple separate networks and reducing overall system complexity.
Data Source
AI summary
An embodiment takes the form of a prediction system that obtains a respective indicated similarity of trajectories in each of a plurality of combinations of observed vehicle trajectories. The prediction system, for each of the combinations, calculates a distance between respective intermediate representations, generated by a neural network, of the trajectories in the combination, and performs a training of the neural network based on a calculated loss between the distance and the indicated similarity of the trajectories in the combination. The respective calculated losses converge after performing respective trainings for the combinations. The prediction system creates an intermediate space of respective intermediate representations, generated by the neural network, of observed vehicle trajectories, and predicts a trajectory of a road agent based on a distance between a given intermediate representation in the intermediate space and an intermediate representation of an observed trajectory of the road agent generated by the neural network.