Road Obstacle Detection by Drivable Area Inversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection models in autonomous vehicles struggle to detect generic obstacles on a road surface, as they are typically trained to recognize predefined classes and fail when encountering unknown objects, leading to inefficiencies and increased costs due to the need for manual annotation and extensive data collection.
Innovation Solution
The approach reframes the problem as occupied/occluded drivable area detection, where a model learns what is and is not on a road surface, using deep learning techniques to differentiate between identifiable and unidentifiable objects, and generates a generic obstacle detection model by retraining a drivable area model with synthetic training images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a conventional object detection model is trained to detect predefined classes of objects, then the model can reliably identify known objects such as vehicles and lane lines, but it fails to detect or correctly classify generic obstacles that do not belong to any predefined class
Solution Approach 1:
Instead of training the model to detect what objects are present (conventional approach), the patent inverts the problem by training the model to detect what is NOT on the road surface. The drivable area model learns to identify the road surface and everything else is considered an obstacle, whether identifiable or generic. This inversion allows the model to detect any object on the road surface without requiring predefined classes or manual annotation of each object type.
2Adaptability or versatility
If manual annotation is performed to add new object classes to the detection model, then the model can detect previously unknown objects, but the process becomes costly and time-consuming
Solution Approach 1:
The system eliminates the need for manual annotation by using synthetic data generation. The drivable area model is trained on synthetically generated images where obstacles are automatically created and labeled through computer graphics techniques. This self-service approach generates unlimited training data without human intervention, allowing the model to learn to detect generic obstacles without any manual annotation of real-world images.
Solution Approach 2:
The patent uses synthetic copying of road surface images with digitally inserted obstacles to create training data. Instead of manually annotating real images, the system creates realistic synthetic copies where obstacles are programmatically added with known ground truth labels, providing unlimited diverse training examples without human annotation effort.
3Measurement precision
If extensive data collection is performed for each new object class, then the model can improve detection accuracy for that class, but the process becomes impractical and impossible for every object on the road
Solution Approach 1:
The drivable area model serves multiple functions simultaneously: it detects the road surface, identifies drivable areas, and flags anything on the road surface as a potential obstacle regardless of type. This universal approach eliminates the need for separate data collection for each object class, as the single model handles all detection tasks through its learned understanding of what constitutes the road surface versus obstacles.
Data Source
AI summary
This application is directed to generic obstacle detection for at least partially autonomous vehicle driving. A first vehicle obtains a road image. The road image includes a road surface along which the first vehicle is travelling. The first vehicle identifies one or more identifiable objects on the road surface in the road image. The first vehicle detects a plurality of objects on the road surface in the road image. The first vehicle eliminates the one or more identifiable objects from the plurality of objects in the road image to determine one or more unidentifiable objects on the road surface in the road image. The first vehicle at least partially autonomously drives the first vehicle by treating the one or more unidentifiable objects differently from the one or more identifiable objects.