Regression-Based Line Detection With Spatial Encoding for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural network (DNN) systems for lane or line detection suffer from a loss of critical spatial information due to down-sampling, leading to reduced precision and accuracy in lane or line detection. Additionally, these systems require significant processing power and resources for real-time deployment.
Innovation Solution
The system employs a regression-based approach that preserves rich spatial information by training a machine learning model to predict distances from each pixel to a line pixel at an input resolution, even at a lower output resolution. This approach reduces runtime by processing inputs at a higher resolution for prediction, while maintaining accuracy through spatial information preservation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional DNN systems use 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 encoding of ground truth data to create target data that captures spatial relationships before the DNN processing. This encoding includes creating distance maps and angle maps that preserve spatial information in a compressed form, allowing the network to work with lower resolution data while maintaining precision through the encoded spatial relationships.
Solution Approach 2:
The patent introduces an intermediary encoding step that transforms the ground truth data into a compact representation that bridges the gap between high-resolution input and low-resolution output. The encoded target data acts as a mediator that contains all necessary spatial information in a form suitable for low-resolution processing, enabling the DNN to learn from compressed representations while maintaining detection accuracy.
2Reliability
If conventional DNN systems process each pixel for each output channel, then comprehensive class predictions are achieved, but runtime increases making real-time deployment difficult
Solution Approach 1:
The system extracts only the essential spatial information needed for line detection from the full pixel data. By encoding ground truth data into compact representations that capture spatial relationships, the system removes redundant information while retaining the critical features needed for accurate detection, significantly reducing processing requirements.
Solution Approach 2:
The patent transforms the problem from predicting class labels for each pixel to predicting encoded spatial parameters (distance and angle) at lower resolution. This parameter transformation allows the system to work with compressed representations that require fewer computations while maintaining the ability to reconstruct accurate line predictions through decoding operations.
Data Source
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.