Lane Boundary Detection Using Drivable Area Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for autonomous vehicle navigation struggle to accurately determine lane boundaries in environments with faded or missing lane markings, leading to inaccurate route planning due to false positives from duplicate or irrelevant markings.
Innovation Solution
A system utilizing two machine learning models, one for detecting drivable areas and another for generating lane boundary information, combines image data to identify drivable areas and outer lane markings, and uses a combined data item to determine inner lane boundaries, even in environments with faded or occluded markings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning models are used to detect lane boundaries, then the system can process standard road conditions, but the accuracy deteriorates in environments with faded or missing lane markings due to false positives from duplicate or irrelevant markings
Solution Approach 1:
The patent divides the lane detection task into two separate segmentation models: one for detecting drivable areas and another for detecting lane boundaries. This segmentation allows each model to specialize in its specific detection task, improving overall accuracy. The drivable area model identifies regions where vehicles can travel, while the lane boundary model specifically identifies lane markings within those regions, reducing false positives from irrelevant markings.
Solution Approach 2:
The patent merges the outputs of two separate machine learning models (drivable area detection and lane boundary detection) into a unified lane identification system. By combining the drivable area mask with the lane boundary detection results, the system achieves more reliable lane identification in challenging environments where either model alone would be insufficient.
2Measurement precision
If a single machine learning model is used for both drivable area and lane boundary detection, then the device complexity is reduced, but the measurement precision deteriorates due to the difficulty of detecting and measuring both features accurately
Solution Approach 1:
The patent segments the detection system into two specialized machine learning models: a drivable area detection model and a lane boundary detection model. Each model is optimized for its specific detection task, allowing both to achieve high measurement precision. The drivable area model focuses on identifying navigable regions, while the lane boundary model focuses on detecting lane markings within those regions.
Solution Approach 2:
The patent creates a unified system that combines multiple detection functions. The lane boundary detection model receives both the original image and the drivable area mask as inputs, allowing it to perform lane boundary detection within the context of known drivable regions. This multi-functional approach maintains precision while managing complexity through shared processing infrastructure.
Data Source
AI summary
Methods, systems, and non-transitory computer-readable media are configured to perform operations comprising determining an image of an environment. A drivable area and boundary information associated with the environment are determined based on the image. At least one boundary for navigation of the environment is generated based on the drivable area and boundary information.