Autonomous Robot Navigation Using Semantic CAD-Based Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional environmental modeling algorithms for autonomous robot navigation face challenges in accuracy and speed, particularly in dense environments, due to limited input data and reliance on 2D feature matching, which results in errors and inefficiencies in localization and mapping.
Innovation Solution
A dense and semantic mapping system using prebuilt CAD models for object recognition and pose estimation, employing deep learning-based object segmentation and adversarial neural networks for accurate and fast object identification and localization, with iterative refinement and Kalman Filter-based optimization for improved accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional dense modeling algorithms are used for environmental mapping, then high-resolution dense mapping is achieved, but execution speed becomes slow and repeated processing is required
Solution Approach 1:
The system performs preliminary action by capturing images at multiple exposure settings (different exposure times or gains) before the actual dense modeling process. These pre-captured images with varying exposure levels are then used to generate the dense map, eliminating the need for repeated processing of the same scene while maintaining high resolution.
Solution Approach 2:
The system applies dynamics by utilizing images captured at different exposure settings dynamically. Instead of processing a single static image multiple times, the algorithm dynamically combines information from multiple exposure-level images to construct the dense map, improving both speed and resolution.
2Ease of operation
If conventional 2D feature matching techniques are used for object recognition, then object identification is achieved, but triangulation accuracy and localization precision are reduced
Solution Approach 1:
The system transitions from 2D feature matching to 3D geometric constraint-based matching. By using known 3D object models and their projections across multiple images with different exposure settings, the system establishes three-dimensional geometric relationships that significantly improve triangulation accuracy and localization precision compared to traditional 2D approaches.
Data Source
AI summary
A method of robot autonomous navigation includes capturing an image of the environment, segmenting the captured image to identify one or more foreground objects and one or more background objects, determining a match between one or more of the foreground objects to one or more predefined image files, estimating an object pose for the one or more foreground objects by implementing an iterative estimation loop, determining a robot pose estimate by applying a robot-centric environmental model to the object pose estimate by implementing an iterative refinement loop, associating semantic labels to the matched foreground object, compiling a semantic map containing the semantic labels and segmented object image pose, and providing localization information to the robot based on the semantic map and the robot pose estimate. A system and a non-transitory computer-readable medium are also disclosed.