Ego Vehicle Behavior Planning With Destination-Weighted Scene Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph-based prediction and planning methods for autonomous vehicles do not effectively consider a specified navigation destination, limiting the ability to generate trajectories that are both safe and compliant with traffic regulations.
Innovation Solution
Modify transition probabilities in the graph-based scene representation to condition the prediction towards a specified navigation destination by increasing the probabilities of expedient routes and decreasing those of non-expedient routes, using pre-trained neural networks for graph encoder, policy header, and trajectory decoder.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If graph-based prediction methods are used without destination conditioning, then general trajectory prediction capability is maintained, but the ability to prioritize navigation-efficient trajectories is lost
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing transition probabilities for all possible graph edges before actual trajectory prediction. This pre-computation conditions the prediction model on the navigation destination in advance, allowing the system to efficiently prioritize destination-oriented trajectories during runtime without adding complex real-time computation requirements.
Solution Approach 2:
The patent introduces transition probabilities as an intermediary element that mediates between the graph-based scene representation and the trajectory prediction output. These transition probabilities serve as a bridge that encodes destination-oriented navigation preferences, allowing the prediction system to favor trajectories that align with the navigation destination without fundamentally changing the prediction architecture.
2Manufacturing precision
If transition probabilities are modified to prioritize navigation destination, then trajectory alignment with destination is improved, but prediction accuracy for non-expedient trajectories may deteriorate
Solution Approach 1:
The patent applies local quality by differentiating transition probabilities based on local graph edge properties relative to the navigation destination. Instead of uniformly modifying all trajectory predictions, the system selectively adjusts transition probabilities for specific graph edges based on their contribution to reaching the destination, thereby improving trajectory precision for navigation-relevant paths while maintaining prediction reliability for other scenarios.
3Ease of manufacture
If pre-trained neural networks are used for graph encoder and trajectory decoder, then model development time is reduced, but adaptability to destination-specific navigation is limited
Solution Approach 1:
The patent applies parameter changes by modifying the transition probability parameters in the pre-trained model based on the navigation destination. Instead of retraining the entire neural network for each destination, the system adjusts the transition probability parameters to reflect destination-oriented navigation preferences, thereby maintaining ease of model development while achieving destination-specific adaptability through parameter adjustment rather than structural changes.
Data Source
AI summary
A computer-implemented method for behavior planning of an at least partially automated ego vehicle having a specified navigation destination. The method includes: generating a graph-based scene representation of the current traffic scene, wherein the graph of the scene representation is defined by a set of nodes and directed edges, each directed edge connecting exactly two nodes, wherein the nodes and/or the edges are associated with map information, and wherein each directed edge is associated with a transition probability for a correspondingly directed transition of the ego vehicle between the connected nodes; planning driving maneuvers, taking into account a future development of the traffic scene. At least a portion of the generated transition probabilities of the scene representation is modified with respect to the specified navigation destination, and that the scene representation having the modified transition probabilities is used for the planning.