Hierarchical Trajectory Anchoring for Nuanced Autonomous Vehicle Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional trajectory prediction models for autonomous vehicles often fail to capture nuanced trajectory predictions, leading to potential severe outcomes, as they typically output only the most likely outcomes and miss complex maneuvers like right turns followed by lane changes to avoid obstacles.
Innovation Solution
The implementation of hierarchical mode anchoring, which breaks down long-horizon trajectories into multiple short-horizon segments, allowing for more nuanced predictions by using a hierarchy of modes such as left-lane changes, right-lane changes, turns, and staying stationary, and representing these as cross products or joint products of segment-level modes, passed as one-hot vectors into prediction models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional trajectory prediction models output only the most likely outcomes, then the prediction model is simple and computationally efficient, but the prediction accuracy and comprehensiveness deteriorate, missing complex maneuvers like right turns followed by lane changes
Solution Approach 1:
The patent segments long-horizon trajectory prediction into multiple short-horizon segments, each predicted independently. This allows the model to capture nuanced maneuvers in each segment while maintaining computational efficiency. The segmentation transforms a single complex prediction task into multiple simpler tasks that can be combined to form comprehensive trajectory predictions.
Solution Approach 2:
The patent introduces a hierarchical dimension to trajectory prediction by organizing predictions at multiple levels: individual trajectory points, short-horizon segments, and long-horizon sequences. This dimensional organization allows the model to output multiple plausible trajectories with different probabilities, improving comprehensiveness without requiring a single overly complex model.
2Loss of information
If hierarchical mode anchoring breaks down trajectories into multiple short-horizon segments, then the prediction comprehensiveness improves, but the computational complexity and processing time increase
Solution Approach 1:
By dividing the prediction horizon into multiple short segments, the model can process each segment independently and in parallel, reducing the overall computational burden compared to processing a single long-horizon prediction. The segmentation enables efficient reuse of computational results across segments.
Solution Approach 2:
The patent performs preliminary anchoring of mode sequences before full trajectory generation. This preliminary action constrains the search space for subsequent trajectory prediction, reducing the computational time required to generate comprehensive predictions while maintaining prediction quality.
3Loss of information
If multiple predicted trajectories are identified and recorded, then the understanding of object movements becomes more nuanced, but the data processing and storage requirements increase
Solution Approach 1:
The patent extracts and anchors only the essential mode sequences from full trajectory predictions. By separating the core mode information from detailed trajectory points, the system maintains comprehensive movement understanding while reducing data volume. Only the critical mode annotations are stored and processed further.
Solution Approach 2:
Instead of generating all possible trajectories and then filtering, the patent inverts the approach by first anchoring mode sequences and then generating trajectories consistent with those modes. This inversion dramatically reduces the number of trajectories that need to be processed and stored, as only trajectories matching the anchored modes are considered.
Data Source
AI summary
The subject disclosure relates to techniques for increasing a quality of predicted trajectories output from a trained prediction algorithm. A process of the disclosed technology can include receiving by a trained prediction algorithm, information about objects in an environment as perceived by sensors of an autonomous vehicle, receiving by the trained prediction algorithm, information about a location of the autonomous vehicle in the environment, generating by the trained prediction algorithm, a predicted trajectory for an object among the objects in the environment, wherein the predicted trajectory being anchored by a path through a tree of paths including at least two modes, wherein each mode creates a node in the tree, wherein a mode is a semantic classification of a portion of the predicted trajectory, and outputting by the trained prediction algorithm, the predicted trajectory for the object.