Large Scale Dense Mapping Using GPU Memory and TSDF
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current large-scale mapping techniques struggle with dense mapping of environments using 2.5D dense depth images and camera matrices, which are not suitable for three-dimensional range data from sensors like LIDAR, leading to memory inefficiencies and processing challenges.
Innovation Solution
The implementation of a large scale dense mapping technique using range data from sensors like LIDAR, which employs a truncated signed distance function (TSDF) integration method to reduce data storage and utilizes GPU memory for efficient processing, dividing the environment into cells and voxels, and swapping data in and out of memory as the vehicle moves, without relying on camera matrices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 2.5D dense depth images and camera matrices are used for mapping, then dense mapping can be achieved, but memory usage increases and processing becomes inefficient for large-scale environments
Solution Approach 1:
The patent segments the environment into discrete cells arranged in a grid structure, where each cell contains only the range data relevant to that specific spatial region. This segmentation allows the system to process and store only necessary portions of the environment at any given time, reducing overall memory requirements while maintaining dense mapping capability within each cell.
Solution Approach 2:
The patent applies local quality by using different data representations for different spatial regions. Each cell stores range data in a localized coordinate system optimized for that region, rather than using a single global camera matrix for the entire environment. This allows efficient memory utilization by storing only locally relevant information with appropriate precision.
2Measurement precision
If 2.5D dense depth images and camera matrices are used for mapping, then dense mapping can be achieved, but processing speed decreases for large-scale environments
Solution Approach 1:
By segmenting the environment into cells, the patent enables parallel processing of multiple cells simultaneously on the GPU. Each cell can be processed independently, allowing the system to scale processing efficiency with large-scale environments without the computational burden of a single global camera matrix transformation for all points.
Solution Approach 2:
The patent transitions from 2.5D depth image representation to a 3D cell-based volumetric representation. This dimensional change allows range data to be organized in three-dimensional space, enabling more efficient spatial queries and reducing the computational complexity of transforming and rendering large-scale environments.
3Area of stationary object
If point based approaches are used for large scale mapping, then mapping of large areas is simplified, but the world is captured sparsely leading to complicated strategies for generating maps
Solution Approach 1:
The patent uses cell segmentation to bridge the gap between sparse point-based approaches and dense mapping. Each cell can be populated with range data at appropriate density, allowing the system to manage large areas efficiently while maintaining dense representation within each cell boundary.
Data Source
AI summary
A large scale dense mapping technique utilizes range data that identifies points in three dimensions. The large scale dense mapping technique utilizes the range data obtained from range sensors coupled with a depth integration technique. After dividing the world, or a portion of the world, into cells and voxels, a mapping system collects range data for a portion of the environment. To speed up the processing involving the large number range measurements, Graphics Processing Unit (GPU) memory, can be used to store data relating to the mapping of the environment. Instead of storing all of the cells of volumetric representation in the memory, just the portion of the cells associated with the range data are stored within the memory. The mapping system can use an SDF or a TSDF based volumetric representation format to represent the environment. The SDF or TSDF based volumetric representation is updated using the range data.