3D Voxel Mapping for Autonomous Positioning Without Object Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for creating high-definition mapping data for autonomous entities, such as vehicles and robots, rely heavily on complex object recognition from point-cloud data generated by LiDAR sensors, which is costly and requires frequent retraining for different environments and sensor manufacturers, and struggle with sparse object recognition in limited environments.
Innovation Solution
A method and system that generate a three-dimensional virtual construct by splitting spatial volumes into voxels, determining voxel occupancy status using LiDAR data, and overlaying visual semantic data from cameras to create a high-definition map, reducing the complexity of object recognition and enabling real-time mapping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex object recognition from point-cloud data is used for autonomous positioning, then positioning accuracy is improved, but computational complexity and cost increase significantly
Solution Approach 1:
The patent segments the continuous point-cloud data into discrete voxel grid cells, transforming the complex object recognition problem into simpler occupancy detection. Each voxel is independently evaluated for occupancy status, dividing the computational task into manageable units that can be processed efficiently.
Solution Approach 2:
The patent creates a simplified voxel-based representation (copy) of the environment that mirrors the essential spatial structure without requiring complex object recognition. This voxel map serves as a computationally efficient surrogate for detailed point-cloud analysis, maintaining positioning accuracy while reducing computational burden.
2Measurement precision
If AI models are retrained for each new environment and sensor manufacturer, then object recognition accuracy is improved, but time and resource consumption increase
Solution Approach 1:
The patent changes the fundamental parameter from object-specific recognition to generic occupancy detection. By evaluating whether voxels are occupied rather than identifying specific objects, the system eliminates the need for environment-specific or manufacturer-specific AI model retraining, while still achieving accurate positioning through spatial pattern matching.
3Measurement precision
If conventional object recognition techniques are used in environments with limited objects, then positioning can be achieved, but the sparsity of objects beyond sensor reach makes localization challenging
Solution Approach 1:
The patent transitions from 2D object detection to 3D voxel-based spatial representation. By utilizing the third dimension and creating a volumetric occupancy map, the system can localize entities even when traditional 2D object features are sparse or unavailable, as the voxel grid captures spatial structure at multiple scales and resolutions.
Data Source
AI summary
A method and system for positioning of autonomously operating entities are disclosed. A positioning system receives a current location of an entity capable of autonomous operation and generates a 3D virtual construct by splitting a spatial volume associated with the current location into a plurality of voxels. The positioning system receives spatial data corresponding to the current location generated by at least one sensor associated with the entity and determines an occupancy status of one or more voxels using the spatial data. A voxel map is configured from the 3D virtual construct based on the occupancy status of the one or more voxels. The positioning system generates a 3D map by overlaying visual semantic data onto the voxel map. The visual semantic data is derived from image frames corresponding to the current location captured by one or more imaging devices. The 3D map is capable of autonomously positioning the entity.