Vehicle Motion Forecasting With Graph-Based Path Conditioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing motion forecasting methods for autonomous vehicles rely on rasterization, which requires a large number of parameters and is inefficient for real-time planning, particularly in complex urban environments where topological graph representations are more suitable for capturing social interactions and lane-following behaviors.
Innovation Solution
A method using a Graph Attention Network (GAT) and a multi-modal decoder to generate diverse predictions without rasterization, incorporating a road-network attention module and dynamic interaction graph to capture geometric and social relationships, allowing for counterfactual reasoning and efficient trajectory planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rasterization is used to represent scene context, then learned models can predict trajectories from extracted features, but the models require a relatively large number of parameters and are inefficient for real-time planning
Solution Approach 1:
The patent replaces the mechanical rasterization process with a graph-based representation system. Instead of converting scene context into pixel-based images that require convolutional neural networks, the system uses topological graphs where nodes represent actors and edges represent relationships. This substitution eliminates the need for rasterization while maintaining trajectory prediction capability, directly addressing the contradiction between prediction accuracy and model complexity.
Solution Approach 2:
The patent changes the fundamental parameters of scene representation from pixel-based raster data to graph-based topological data. By transforming the representation format from continuous pixel grids to discrete graph structures with nodes and edges, the system reduces the number of parameters required while preserving essential spatial and relational information for accurate trajectory prediction.
2Measurement precision
If rasterization is used to represent scene context, then trajectory prediction can be performed, but it is inefficient for real-time planning and interactive probing of futures
Solution Approach 1:
The patent substitutes the computationally intensive rasterization and convolutional processing with efficient graph-based operations. The graph representation allows for direct computation of trajectories through message passing and attention mechanisms, eliminating the need for repeated rasterization and convolution operations, thereby enabling real-time planning and interactive probing of multiple future scenarios.
Solution Approach 2:
The patent introduces dynamic graph structures that can be efficiently updated and queried in real-time. The graph representation allows the system to dynamically adjust to new information and recompute trajectories quickly, supporting interactive planning where multiple future scenarios can be explored without re-processing the entire scene context from scratch.
3Adaptability or versatility
If rasterization is used, then some facets of the problem can be represented, but coordinate spaces not conducive to rasterization (such as curvilinear coordinates for lane-following) cannot be easily represented
Solution Approach 1:
The patent changes the coordinate representation paradigm from fixed pixel grids to flexible graph-based coordinate systems. This allows the system to naturally represent curvilinear coordinates for lane-following behaviors and other non-Cartesian coordinate systems by defining nodes and edges in the appropriate coordinate space, eliminating the need to force all data into raster-compatible formats.
Solution Approach 2:
The patent creates a universal graph-based representation framework that can handle multiple coordinate systems and representation types within a single unified structure. The graph model can represent actors, lanes, intersections, and other road features in their native coordinate systems while maintaining consistent relational semantics, providing versatility across different spatial representations without requiring separate processing pipelines.
4Measurement precision
If rasterization is used to represent scene context, then trajectory prediction can be performed, but the approach cannot easily provide top-down feedback and causal explainability required for safety critical applications
Solution Approach 1:
The patent replaces the opaque rasterization and convolutional processing with interpretable graph-based operations. The graph structure explicitly represents causal relationships between actors, lanes, and traffic elements, allowing the system to provide top-down feedback and explain predictions in terms of the underlying graph structure and relationships, thereby maintaining accuracy while improving interpretability for safety-critical applications.
Data Source
AI summary
Systems and methods of determining trajectories of an actor in an environment in which a vehicle is operating are provided. The method includes detecting an actor that may move within a scene in the environment by an object detection system of a vehicle in the environment, determining a kinematic history of the actor, and using context of the scene and the kinematic history of the actor to determine a plurality of reference polylines for the actor. The method further includes generating a contextual embedding of the kinematic history of the actor to generate a plurality of predicted trajectories of the actor, in which the generating conditions each of the predicted trajectories to correspond to one of the reference polylines. The method additionally includes using, by the vehicle, the plurality of predicted trajectories to plan movement of the vehicle.