Image-Based Lane Index Estimation for Weak-Signal Vehicle Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing localization methods for autonomous vehicles rely on GNSS and IMU, which can be computationally expensive or inaccurate in areas with reduced signal reception, leading to errors in location and resolution.
Innovation Solution
Utilizing machine learning models trained on historical image data to generate lane offset and index information based on real-time image data from sensors, such as LiDAR and cameras, to improve localization accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing localization methods use GNSS and IMU, then location information can be obtained, but computational expense increases and accuracy deteriorates in areas with reduced signal reception
Solution Approach 1:
The patent replaces the mechanical/GNSS-based localization system with an image-based machine learning system. Instead of relying on satellite signals and inertial measurements, the system uses image data from cameras and LiDAR sensors processed through trained machine learning models to determine lane offset and localization information, thereby avoiding computational expenses and signal reception issues associated with traditional methods
Solution Approach 2:
The patent introduces machine learning models as an intermediary between raw image data and localization output. The trained models act as mediators that translate image features directly into lane offset and position information, eliminating the need for computationally intensive processing of GNSS and IMU data while maintaining or improving accuracy
2Measurement precision
If existing localization methods use GNSS and IMU, then location data can be obtained, but location resolution deteriorates in areas with reduced signal reception
Solution Approach 1:
The patent substitutes the GNSS/IMU mechanical localization system with an optical/image-based system using machine learning. This replacement enables high-resolution localization based on visual features from cameras and LiDAR, achieving consistent measurement precision regardless of satellite signal availability
Solution Approach 2:
The machine learning models are trained on historical image data and lane offset data to learn associations independently, enabling the system to self-determine localization without external signal assistance. The models use learned patterns from training data to accurately predict lane offset and position purely from image features
3Measurement precision
If machine learning models are trained on historical image data to generate lane offset data, then localization accuracy improves, but device complexity increases
Solution Approach 1:
The patent applies preliminary action by training machine learning models offline using historical image data and lane offset data before deployment. The models learn associations between image features and lane offset during the training phase, so that during actual operation, they can quickly and accurately predict localization from image data without complex real-time processing
Solution Approach 2:
The patent transitions from traditional coordinate-based localization to a dimension-based lane index system. Instead of directly computing continuous position coordinates, the system uses machine learning models to predict discrete lane indices and offsets, simplifying the computational dimension while improving localization accuracy in multi-lane roadways
Data Source
AI summary
Systems and methods for training and executing machine learning models to generate lane index values are disclosed. A method includes identifying a set of image data captured by at least one autonomous vehicle when the at least autonomous vehicle is positioned in a lane of a roadway and respective ground truth localization data; determining a plurality of lane index values for the set of image data based on the ground truth localization data; labeling the set of image data with the plurality of lane index values, the lane index values representing a number of lanes from a leftmost or rightmost lane to the lane in which the at least one autonomous vehicle was positioned; and training, using the labeled set of image data, a plurality of machine learning models that generate a left lane index value and a right lane index value as output.