Mobile Robot Traversability Evaluation With Self-Trained Elevation Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current mobile robots face challenges in navigating urban environments due to the sensitivity of vision-based methods to illumination changes and the limitations of 2D LiDAR sensors in detecting 3D terrain obstacles, leading to potential hazards like steps, stairs, and dynamic obstacles, which necessitate a more reliable traversability analysis for safe and efficient navigation.
Innovation Solution
A mobile robot system that generates a grid-cell based elevation map using point cloud data, extracts features such as absolute height difference, slope, curvature, and edge, and employs a self-training unit to create an AI model for traversability evaluation by iteratively labeling and retraining on labeled and unlabeled data sets, classifying grid cells as traversable or non-traversable based on predefined thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If vision-based methods are used for traversability analysis, then dense point cloud information is provided, but sensitivity to illumination changes occurs
Solution Approach 1:
The patent combines 2D LiDAR data with camera images to create a fused point cloud system. The 2D LiDAR provides accurate depth information while the camera provides visual context, merging the strengths of both sensors to achieve illumination-resistant traversability analysis with dense point cloud coverage.
Solution Approach 2:
The patent introduces an intermediary processing system that fuses 2D LiDAR point cloud data with camera imagery. This intermediary fusion layer allows the system to leverage the illumination independence of LiDAR while incorporating the dense visual information from the camera, resolving the contradiction between point cloud density and illumination sensitivity.
2Ease of operation
If 2D LiDAR sensors are used for navigation, then autonomous navigation in structured environments is achieved, but detection of 3D terrain obstacles is limited
Solution Approach 1:
The patent transitions from 2D LiDAR scanning to 3D point cloud reconstruction by integrating 2D LiDAR data with camera imagery. This dimensional enhancement allows the system to detect 3D terrain obstacles like steps and stairs while maintaining the autonomous navigation capabilities enabled by structured 2D scanning.
3Measurement precision
If supervised learning with hand-labeled data is used for traversability analysis, then classification accuracy is improved, but training data preparation complexity increases
Solution Approach 1:
The patent implements self-training mechanisms where the system automatically generates training data from its own operational experiences and unlabeled point cloud data. The system performs self-labeling and iterative retraining, eliminating the need for manual data annotation while maintaining high classification accuracy through autonomous learning.
Data Source
AI summary
The present invention relates to a mobile robot for evaluating self-training based traversability comprising: an elevation map generator which generates a grid-cell based elevation map using point cloud data; a feature extractor which extracts a plurality of types of features on each grid cell from the elevation map; a data set generator which generates a labeled data set labeled for training and an unlabeled data set, based on label features set for at least two types of features among the plurality of types of features; and a self-training unit which generates an AI model for evaluating traversability by self-training using the labeled data set and the unlabeled data set. Accordingly, it is possible to create training data which increases training ability and then use the data for the self-training, whereby traversability can be evaluated while achieving navigation safety and efficiency.