Drivable Free-Space Boundary Detection for Real-Time Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous driving systems face challenges in determining drivable free-space in real-time due to computationally expensive vision-based techniques, such as deep neural networks (DNNs), which often compromise on accuracy or require excessive computing resources, failing to account for dynamic obstacles and boundary classes.
Innovation Solution
A fully convolutional neural network (CNN) is employed to regress on boundaries separating drivable free-space from non-drivable space, eliminating fully connected layers to reduce computational demands and incorporating contextual information about boundary classes, enabling accurate and efficient real-time operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If semantic segmentation using deep neural networks is used to determine drivable free-space, then classification accuracy is improved, but computational expense increases making real-time operation impossible
Solution Approach 1:
The patent segments the image processing task by treating each column independently rather than performing full semantic segmentation on all pixels. This column-wise regression approach processes one vertical slice at a time, dramatically reducing computational complexity while maintaining boundary detection accuracy for drivable free-space determination.
Solution Approach 2:
The patent extracts only the essential information needed for autonomous driving by directly regressing boundary locations and classes from image columns, rather than classifying every pixel. This extraction of critical boundary data eliminates unnecessary computational overhead while preserving the information needed for safe navigation.
2Use of energy by moving object
If conventional CNNs with fully connected layers are used for column-wise regression, then computational resources are reduced, but the type or class of boundary is not identified reducing safety
Solution Approach 1:
The patent merges two functions into a single neural network output: boundary location regression and boundary class classification. The CNN simultaneously predicts both the position of drivable free-space boundaries and their semantic classes (e.g., curb, pedestrian, vehicle), ensuring comprehensive safety information is captured without requiring separate processing stages.
Solution Approach 2:
The patent adds a classification dimension to the regression output by having the neural network predict both continuous boundary positions and discrete boundary classes. This multi-dimensional output provides both spatial precision and semantic context, enabling the autonomous vehicle to understand not just where boundaries are but what they represent for safe navigation.
Data Source
AI summary
In various examples, sensor data may be received that represents a field of view of a sensor of a vehicle located in a physical environment. The sensor data may be applied to a machine learning model that computes both a set of boundary points that correspond to a boundary dividing drivable free-space from non-drivable space in the physical environment and class labels for boundary points of the set of boundary points that correspond to the boundary. Locations within the physical environment may be determined from the set of boundary points represented by the sensor data, and the vehicle may be controlled through the physical environment within the drivable free-space using the locations and the class labels.