Autonomous Route Training Data from Odometry and Obstacle Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for autonomous driving in complex urban environments are limited by the need for large-scale manually-annotated datasets, which are time-consuming and expensive, and struggle with occlusions, shadows, and changing lighting conditions, especially in areas with few or no lane markings.
Innovation Solution
A weakly-supervised approach that uses vehicle odometry and obstacle sensing data to label images without manual annotation, training a segmentation unit to identify drivable paths using a data collection vehicle's actual path and obstacle detection, enabling the generation of vast quantities of labeled training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods use pre-processing steps to remove shadow and exposure artefacts, then road feature extraction is improved, but the system fails in the presence of occlusions, shadows and changing lighting conditions
Solution Approach 1:
The patent uses LIDAR data to create a 3D copy of the road surface geometry and projects it onto the 2D image plane to generate a synthetic depth map. This copy provides geometric information that is independent of lighting conditions, allowing the system to distinguish between actual road features and artifacts caused by shadows or exposure variations.
Solution Approach 2:
The system combines multiple data sources (color image, depth map from LIDAR, and geometric model) into a composite representation of the road scene. By fusing these different types of information, the system achieves both color-based semantic understanding and geometrically-accurate road boundary detection that is robust to lighting variations.
2Reliability
If radar or LIDAR is combined with images to increase robustness, then reliability is improved, but sensor cost increases
Solution Approach 1:
The LIDAR system performs multiple functions: it provides both obstacle detection data and geometric road surface information. The same sensor that detects obstacles also generates the depth map used to correct road boundary detection, eliminating the need for separate sensors and reducing overall system cost.
Solution Approach 2:
The patent introduces a road surface geometric model as an intermediary that bridges the gap between 2D image data and 3D road geometry. This model acts as a mediator that translates LIDAR depth information into corrected road boundaries, allowing the system to achieve LIDAR-level robustness with more affordable sensing configurations.
3Reliability
If deep learning networks are used for semantic segmentation, then robustness to lighting changes is improved, but large-scale manually-annotated datasets are required which are time-consuming and expensive to produce
Solution Approach 1:
The system uses the vehicle's own LIDAR data and odometry information to automatically generate training labels for the deep learning network. The vehicle's recorded trajectory and obstacle detections serve as self-generated supervision signals, eliminating the need for manual annotation while providing accurate ground truth data for training.
Solution Approach 2:
The patent performs preliminary road surface modeling and geometric correction during the data collection phase. By pre-processing the LIDAR data to create accurate depth maps and road surface models before training the network, the system prepares high-quality training data that reduces the amount of manual annotation needed and speeds up the overall training process.
4Productivity
If feature-based methods are used for lane detection, then computational efficiency is improved, but performance degrades in areas with few or no lane markings
Solution Approach 1:
The patent transitions from 2D image-based feature detection to 3D geometric reasoning by incorporating LIDAR depth information. By lifting the problem into three dimensions and using the known road surface geometry, the system can detect road boundaries even in the complete absence of 2D visual features like lane markings, while maintaining computational efficiency through geometric calculations.
Data Source
Figure 1A
Figure 1B
Figure 1B
AI summary
A method (1300) of generating a training dataset for use in autonomous route determination, the method comprising obtaining (1302) data from a data collection vehicle (10) driven through an environment. The data comprises vehicle odometry data detailing a path taken by the vehicle(10) through the environment,obstacle sensing data detailing obstacles detected in the environment; and images (106) of the environment. The method (1300) further comprises using (1304) the obstacle sensing data to label one or more portions of at least some of the images (108) as obstacles and using (1306) the vehicle odometry data to label one or more portions of at least some of the images (108) as the path taken by the vehicle (10) through the environment.