Autonomous Line Detection Using Pixel-Distance Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional lane and line detection systems for autonomous vehicles suffer from loss of spatial information due to down-sampling in deep neural networks, leading to reduced precision and accuracy, increased processing time, and high computational demands.
Innovation Solution
A regression-based approach that preserves spatial information by training a machine learning model to predict distances and angles at input resolution, using bit encoding and clustering to reduce processing time and computational burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If deep neural networks perform down-sampling during processing, then computational complexity is reduced, but spatial information is lost leading to reduced detection precision
Solution Approach 1:
The system performs preliminary action by predicting pixel distances to line pixels before final line reconstruction. The machine learning model outputs distance values for each pixel indicating how far it is from the nearest line pixel, preserving spatial information even at lower resolutions. This preliminary distance prediction allows accurate line detection without requiring full-resolution processing throughout the entire pipeline.
Solution Approach 2:
The patent introduces an intermediary representation - the pixel distance map - that bridges the gap between low-resolution processing and high-precision detection. Instead of directly processing high-resolution images through the entire DNN pipeline, the system uses distance predictions as an intermediary that encodes spatial relationships at lower resolution while enabling precise line reconstruction afterward.
2Measurement precision
If conventional systems use separate output channels for each line class, then classification accuracy is maintained, but processing time increases
Solution Approach 1:
The patent fundamentally changes the output parameter from class probabilities to pixel distances. Instead of predicting which class each pixel belongs to using multiple output channels, the system predicts a single distance value for each pixel indicating its distance to the nearest line pixel. This parameter transformation reduces the number of output channels needed while maintaining detection accuracy through the distance-based reconstruction approach.
3Loss of information
If high-resolution images are processed throughout the pipeline, then spatial information is preserved, but computational resources and energy consumption increase
Solution Approach 1:
The system performs preliminary distance prediction at lower resolution, then uses these distance values to reconstruct lines at the original high resolution. This allows the computationally intensive DNN processing to operate on down-sampled images while still achieving high-resolution detection results through the distance-based reconstruction method.
Solution Approach 2:
The patent creates a compressed representation (distance map at lower resolution) that captures essential spatial information, then uses this copy to reconstruct the full-resolution line detection results. This copying approach allows the system to work with smaller data structures during processing while maintaining the ability to produce high-resolution output.
4Productivity
If down-sampling is applied to reduce processing burden, then processing speed increases, but line detection accuracy decreases
Solution Approach 1:
The system performs preliminary distance prediction on down-sampled images, then uses these distance values to accurately reconstruct lines at full resolution. The preliminary action of predicting distances at lower resolution captures sufficient spatial information to enable accurate high-resolution line detection without requiring the entire processing pipeline to operate at full resolution.
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
In various examples, systems and methods are disclosed that preserve rich spatial information from an input resolution of a machine learning model to regress on lines in an input image. The machine learning model may be trained to predict, in deployment, distances for each pixel of the input image at an input resolution to a line pixel determined to correspond to a line in the input image. The machine learning model may further be trained to predict angles and label classes of the line. An embedding algorithm may be used to train the machine learning model to predict clusters of line pixels that each correspond to a respective line in the input image. In deployment, the predictions of the machine learning model may be used as an aid for understanding the surrounding environment - e.g., for updating a world model - in a variety of autonomous machine applications.