Autonomous Vehicle Mover Model Selection for Ambiguous Actor Motion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicle (AV) control systems face challenges in predicting the motion of actors when they fall into multiple classifications, leading to less reliable trajectory forecasting.
Innovation Solution
The system determines which kinematic model an AV should use by processing sensor data to identify probable classes of a detected actor, querying a library of kinematic models, applying multiple models to predict trajectories, and selecting the most accurate model based on kinematic states and divergence evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the AV uses a single kinematic model for an actor, then the system complexity is reduced, but the trajectory forecasting reliability deteriorates when the actor belongs to multiple classifications
Solution Approach 1:
The system segments the kinematic model selection process into multiple discrete steps: (1) detecting actor characteristics, (2) assigning multiple probable classes with confidence scores, (3) querying the library for multiple candidate models, (4) evaluating each model's predicted trajectory against detected characteristics, and (5) selecting the model with the lowest divergence. This segmentation allows the system to handle multi-classification scenarios systematically without overwhelming complexity.
Solution Approach 2:
The system changes the selection criterion from simple class probability to trajectory divergence measurement. Instead of directly selecting based on classification confidence, the system evaluates how well each candidate model's predicted trajectory matches the actually detected kinematic characteristics, using this divergence metric to select the most appropriate model.
2Reliability
If the AV queries the library for multiple kinematic models, then the trajectory forecasting reliability is improved, but the computational time and processing complexity increase
Solution Approach 1:
The system performs preliminary actions by maintaining a pre-established library of kinematic models organized by actor class, where each model is already associated with specific actor characteristics. When an actor is detected, the system quickly queries this pre-organized library using the detected characteristics as keys, rather than evaluating all possible models from scratch, thus reducing computational time.
Solution Approach 2:
The system substitutes complex real-time model evaluation with a divergence calculation approach. Instead of running full simulations to compare model predictions, the system calculates a simplified divergence metric between predicted trajectory parameters and actually detected kinematic characteristics, significantly reducing computational requirements while maintaining reliability.
3Ease of operation
If the AV selects the kinematic model based on highest class probability, then the ease of operation is improved, but the manufacturing precision of model selection deteriorates when class boundaries are ambiguous
Solution Approach 1:
The system introduces an intermediary evaluation step between class probability and model selection. Instead of directly selecting the model associated with the highest probability class, the system uses detected kinematic characteristics as an intermediary to evaluate how well each candidate model's predictions align with actual observations, providing a more precise selection criterion that bridges classification and model performance.
Data Source
AI summary
This document discloses system, method, and computer program product embodiments for operating a vehicle, comprising: using kinematic models to generate forecasted trajectories of an actor (the kinematic models being respectively associated with different actor types that are assigned to an actor detected in an environment of the vehicle); selecting a first kinematic model based on the forecasted trajectories and a kinematic state of the actor; using the first kinematic model to predict a first path for the actor; selecting a second kinematic model responsive to movement of the actor no longer being consistent with typical movement of an object of one of the different actor types that is associated with the first kinematic model; using the second kinematic model to predict a second path for the actor; and controlling operations of the vehicle based on the first and second paths.