Vehicle Camera Lane Detection for Curved and Merging Roads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing lane detection systems struggle to accurately detect tightly curved lanes and handle situations involving merging and dividing lanes, often failing in urban environments due to assumptions about lane monotonicity and requiring complex post-processing.
Innovation Solution
A method and device utilizing a convolutional neural network (CNN) with a feature extractor to detect lane candidates, applying a grid over camera images, filtering based on confidence thresholds, transforming keypoints to 3D coordinates, and interpolating additional points for enhanced resolution, enabling detection of tightly curved lanes and merging/dividing scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If existing lane detection systems use assumptions about lane monotonicity, then detection simplicity is improved, but detection accuracy for tightly curved lanes deteriorates
Solution Approach 1:
The patent changes the parameter representation from assuming monotonic lane boundaries to using keypoint coordinates that can represent arbitrary curved lanes. The lane is represented by a sequence of keypoints (x1, y1), (x2, y2), ..., (xn, yn) where each keypoint can independently position itself, allowing the lane to curve tightly without requiring monotonicity assumptions. This resolves the contradiction by changing how lane geometry is parameterized.
2Measurement precision
If complex post-processing is applied to handle merging and dividing lanes, then detection accuracy for complex lane situations is improved, but processing time increases
Solution Approach 1:
The patent segments the lane detection problem into independent keypoint detections within a grid structure. Each cell in the grid independently detects keypoints, and these keypoints are then connected to form lane segments. This segmentation allows complex merging and dividing lanes to be handled by simple local detections that are then assembled, avoiding the need for complex global post-processing while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary action by pre-defining a grid structure over the image space before detection. This grid provides a fixed framework that guides the detection of keypoints throughout the image. By establishing this structure in advance, the system avoids the need for complex real-time post-processing to handle merging and dividing lanes, as the grid naturally accommodates these scenarios through its regular structure.
3Productivity
If a fixed number of keypoints per lane candidate is used, then processing efficiency is improved, but adaptability to different lane configurations deteriorates
Solution Approach 1:
The patent applies dynamics by making the number of keypoints adaptive rather than fixed. The system detects a variable number of keypoints based on the actual lane configuration in the image. For simple straight lanes, fewer keypoints are detected, while for complex curved or merging lanes, more keypoints are automatically detected. This dynamic adaptation maintains processing efficiency while achieving versatility across different lane configurations.
Data Source
AI summary
A method for detecting a lane for a driver assistance system. The method includes an arranging step in which a grid is arranged over a camera image from a vehicle camera, the camera image displaying a roadway. In a displaying step, a lane candidate is displayed using at least one parameter and/or a confidence value for the lane candidate when the at least one lane candidate is detected in at least one cell of the grid. In a filtering-out step, the lane candidate is filtered out when the confidence value of the lane candidate is below a defined confidence threshold. In a back-transforming step, the lane candidate is transformed back to a fixed number of keypoints per lane candidate when the confidence value of the lane candidate is above the confidence threshold, to obtain the lane. In a determining step, 3D coordinates for the lane are determined using the keypoints.