Lane Detector Training Using Marking-Removed Road Images
Solution Overview
Problem
Existing lane detection systems struggle to accurately recognize lanes on unmarked roads without relying on expensive HD maps or manual labeling, leading to reduced reliability and safety in automated driving systems.
Innovation Solution
A method for training a lane detector using a data set created by removing lane markings from digital images and replacing them with surrounding textures, utilizing generative models like stable diffusion or GANs, to enhance the detector's ability to identify lanes in real-world conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If lane markings are removed from training images to simulate unmarked roads, then the lane detector's ability to handle unmarked roads is improved, but the availability of ground truth data for training becomes more difficult
Solution Approach 1:
The patent creates synthetic training images by copying existing marked road images, removing lane markings through generative models, and preserving the corresponding lane geometry data as ground truth. This copying approach allows unlimited generation of unmarked road training samples without requiring additional real-world data collection.
Solution Approach 2:
The patent performs preliminary lane detection and geometry extraction on marked images before removing the markings. By pre-extracting the lane fundamental matrices and geometric parameters from the original marked images, the system ensures ground truth data is available before the markings are removed, solving the chicken-and-egg problem of training data generation.
2Reliability
If generative models are used to remove lane markings and replace with surrounding textures, then the realism of unmarked road images is improved, but the computational complexity and training time increase
Solution Approach 1:
The generative model is trained on the synthetic unmarked road images generated by the system itself, creating a self-reinforcing loop where the model continuously improves its texture synthesis capability. This self-service approach allows the system to generate increasingly realistic unmarked road images without requiring external annotated data, reducing overall training time in the long run.
3Productivity
If manual labeling is replaced by automated lane marking detection, then the productivity of data preparation is improved, but the accuracy of ground truth data may deteriorate
Solution Approach 1:
The system implements feedback loops where the automated lane detection results are evaluated and refined. The generative model's texture synthesis is guided by feedback from the lane geometry data, ensuring that the removed markings are replaced with textures that match the surrounding road surface while preserving the underlying lane structure for accurate ground truth.
Data Source
AI summary
A computer-implemented method is disclosed for training a lane detector for detecting a lane in a digital image, in particular based on pixel values of the digital image. The method includes (i) providing a first plurality of digital images with marked lanes labeled with lane markings, (ii) obtaining a second plurality of digital images from the first plurality of digital images by removing the lane markings from the digital images of the first plurality of digital images, thereby generating corresponding digital images without markings, (iii) providing data for characterizing the lanes in the digital images of the first plurality of digital images, and (iv) providing a data set for training the lane detector. The data set includes pairs of a digital image and data characterizing a fundamental truth of a lane in the digital image. The digital image is taken from the second plurality of digital images, and the data characterizes the lane identified in the corresponding digital image from the first plurality of digital images.



