3D Drivable Space Detection From Vehicle Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current road surface detection methods for autonomous navigation often fail to differentiate between drivable and non-drivable spaces, limiting the accuracy of vehicle path determination.
Innovation Solution
The method generates 3D bounding boxes from 2D bounding boxes of objects in captured images, using geometric constraints to determine non-drivable spaces, which are then used to train machine learning models to classify image portions as drivable or non-drivable, assisting in autonomous navigation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained to predict road surfaces from sensor data, then road surface detection capability is improved, but the ability to determine whether surfaces are actually drivable deteriorates
Solution Approach 1:
The patent transitions from 2D image data to 3D spatial understanding by generating 3D bounding boxes from 2D detections. This dimensional transformation enables the system to infer depth, volume, and spatial relationships, allowing it to determine drivability by analyzing the 3D space occupied by objects and their relationship to the road surface, rather than relying solely on 2D pixel classification.
Solution Approach 2:
The patent introduces 3D bounding boxes as an intermediary representation between 2D object detection and drivability determination. These bounding boxes serve as a bridge that encodes spatial information about objects' positions, sizes, and orientations, which then feeds into the drivability analysis by defining non-drivable regions in 3D space that can be projected back to 2D road surfaces.
2Measurement precision
If 3D bounding boxes are generated from 2D bounding boxes using geometric constraints, then non-drivable space identification is improved, but computational complexity increases
Solution Approach 1:
The patent applies geometric constraints as preliminary conditions before solving for 3D bounding box parameters. By pre-defining relationships between 2D and 3D coordinates, object orientations, and spatial relationships, the system reduces the search space and enables more efficient computation of 3D parameters from 2D inputs, rather than attempting to solve all parameters simultaneously without constraints.
3Reliability
If machine learning models classify image portions as drivable or non-drivable, then navigation safety is improved, but training data requirements increase
Solution Approach 1:
The patent enables the system to generate its own training data by automatically creating labeled drivable/non-drivable regions from 3D bounding boxes and geometric constraints. Instead of requiring manually annotated training datasets, the system uses its own 3D reconstruction and projection capabilities to self-generate supervised learning labels, reducing dependency on external data sources and enabling continuous self-improvement.
Data Source
AI summary
Systems and methods for determining the drivable space of a road, for applications such as autonomous navigation. To determine the non-drivable space under another vehicle, systems and methods of embodiments of the disclosure generate 3D bounding boxes from 2D bounding boxes of objects in captured roadway images, and from various geometric constraints. Image portions may be labeled as drivable or non-drivable according to projections of these 3D bounding boxes onto their road surfaces. These labeled images, along with accompanying semantic information, may be compiled to form training datasets for a machine learning model such as a CNN. The training datasets may train the CNN to classify input image portions into drivable and non-drivable space, for applications such as autonomous navigation.