Adaptive AV Trajectory Prediction With Relevant Object Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in predicting the position and motion of objects in their vicinity quickly and accurately to prevent collisions and make effective decisions, especially in complex environments where not all detected objects are relevant for decision-making.
Innovation Solution
The system determines the relevancy of objects based on a prediction context that includes localization, mode of operation, and driving strategy, using a tree construct to generate relevancy values and cache relevant objects, and selects appropriate trajectory prediction approaches to output predicted trajectories efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If trajectory prediction is performed for all detected objects, then prediction accuracy is improved, but computational load increases
Solution Approach 1:
The system extracts only the relevant subset of objects from the complete set of detected objects based on relevancy criteria such as distance to AV, object type, and motion characteristics. This extraction principle filters out irrelevant objects (e.g., stationary objects far from the AV) before trajectory prediction, thereby reducing computational load while maintaining prediction accuracy for critical objects.
Solution Approach 2:
The system applies different relevancy assessment criteria to different types of objects based on their local characteristics. For example, moving objects closer to the AV are assigned higher relevancy scores and undergo detailed trajectory prediction, while stationary objects or those beyond a threshold distance are excluded. This localized quality assessment optimizes computational resource allocation.
2Measurement precision
If multiple trajectory prediction approaches are used, then prediction accuracy is improved, but processing time increases
Solution Approach 1:
The system applies multiple trajectory prediction approaches selectively rather than uniformly to all objects. For highly relevant objects, multiple approaches (e.g., probabilistic, kinematic, learning-based) are used to generate accurate predictions. For less critical objects, a single simplified approach or no prediction is performed. This partial application of multiple methods balances accuracy requirements with processing time constraints.
3Productivity
If relevancy filtering is applied, then processing efficiency is improved, but risk of missing relevant objects increases
Solution Approach 1:
The system performs preliminary relevancy assessment using readily available object attributes (distance, type, motion state) before committing to full trajectory prediction. This preliminary filtering action uses conservative thresholds to ensure that potentially relevant objects are not prematurely discarded. Objects that meet minimum relevancy criteria are retained for further processing, maintaining reliability while improving efficiency.
Solution Approach 2:
The system incorporates feedback mechanisms where prediction outcomes and object behavior patterns inform future relevancy assessments. If an object initially filtered out subsequently exhibits relevant behavior (e.g., sudden motion), the system can re-evaluate and include it in prediction. This feedback loop ensures that efficiency optimizations do not compromise detection reliability.
Data Source
AI summary
A method may include obtaining one or more inputs in which each of the inputs describes at least one of: a state of an autonomous vehicle (AV) or a state of an object; and identifying a prediction context of the AV based on the inputs. The method may also include determining a relevancy of each object of a plurality of objects to the AV in relation to the prediction context; and outputting a set of relevant objects based on the relevancy determination for each of the plurality of objects. Another method may include obtaining a set of objects designated as relevant to operation of an AV; selecting a trajectory prediction approach for a given object based on context of the AV and characteristics of the given object; predicting a trajectory of the given object using the selected trajectory prediction approach; and outputting the given object and the predicted trajectory.