Vehicle Pose Estimation Using Neural Network Road Marking Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for determining a vehicle's pose, such as those used in autonomous driving for Forward Collision Warning and Automatic Emergency Braking, face challenges in accurately estimating Time-To-Collision and range due to variations in vehicle orientation and road surface conditions, particularly when dealing with curving lane lines or non-parallel road markings, leading to computational inefficiencies and reduced accuracy.
Innovation Solution
A method utilizing a trained artificial neural network to process unordered points associated with road markings, which are pre-processed and normalized, to determine the vehicle's pose efficiently, reducing computational burden and improving accuracy by using a custom neural network architecture that applies non-linear regression.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If classical computer vision approaches are used to determine vehicle pose from road markings, then the system can handle curving lane lines, but the computational complexity increases and accuracy decreases when mathematical assumptions are violated
Solution Approach 1:
The patent transforms the pose estimation problem from direct coordinate regression to parameter-space optimization by representing road markings in polar coordinates (rho, theta) and optimizing these parameters to satisfy geometric constraints. This parameter transformation allows the system to handle curving lane lines while maintaining accuracy by working in a parameter space where the optimization naturally accounts for non-parallel geometries.
Solution Approach 2:
The patent creates a virtual copy of the road marking geometry in parameter space (rho-theta space) rather than directly processing the raw pixel coordinates. This copied representation in an alternative parameter space allows the system to reason about geometric relationships without being constrained by the complexities of direct coordinate transformations, thereby maintaining accuracy for curving lanes.
2Adaptability or versatility
If deep learning methods are used to estimate vehicle pose from single images, then the system can handle complex road geometries, but the computational cost increases significantly
Solution Approach 1:
The patent replaces the mechanical deep learning inference process with a physics-based geometric optimization approach. Instead of using neural networks to learn pose from images, the system uses constraint satisfaction and parameter optimization based on known geometric relationships between road markings and vehicle pose, dramatically reducing computational cost while maintaining versatility.
Solution Approach 2:
The patent introduces polar coordinates (rho, theta) as an intermediary parameter space between the image coordinate system and the vehicle pose parameters. This intermediary representation simplifies the geometric relationships and enables efficient optimization without requiring complex deep learning models, thus improving computational efficiency while handling complex geometries.
3Reliability
If Visual Odometry or SLAM systems are used to determine vehicle pose, then the system can provide continuous pose estimation, but the computational burden increases
Solution Approach 1:
The patent extracts only the essential geometric constraints needed for pose estimation from the full Visual Odometry or SLAM framework. By taking out just the road marking geometric relationships and ignoring the full SLAM complexity, the system achieves continuous pose estimation with significantly reduced computational overhead, maintaining reliability for the specific application of road-based pose determination.
Data Source
AI summary
A set of unordered points associated with road markings is received. The unordered points are inputted to a trained artificial neural network. Using the artificial neural network, an output is generated by applying non-linear regression to the unordered points. Based on the output, a pose of a vehicle is determined.