Lidar Mapping Multithreading Parallel Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current lidar mapping with a moving scanner requires offline batch processing to produce high-quality maps, and real-time mapping algorithms suffer from accuracy penalties due to lower frequency of mapping updates, which can be offset by improving odometry motion estimation but at the cost of increased latency.
Innovation Solution
A multithreading architecture is employed to parallelize computationally expensive mapping operations, allowing for increased frequency of map updates and improved accuracy while reducing reliance on initialization, potentially dispensing with odometry and using IMU measurements for initialization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If offline batch processing is used for lidar mapping, then high-quality maps are produced, but real-time performance is not achieved
Solution Approach 1:
The patent segments the mapping process into multiple threads: odometry thread for motion estimation, mapping thread for point cloud registration, and map management thread for accumulated map updates. This parallelization enables real-time processing while maintaining high map quality by allowing simultaneous execution of computationally intensive operations.
Solution Approach 2:
The system performs preliminary untwisting of point clouds using odometry estimates before mapping registration. This preliminary motion compensation prepares the data for more efficient registration operations, enabling real-time processing without sacrificing mapping accuracy.
2Loss of time
If mapping frequency is increased for real-time performance, then latency is reduced, but accuracy deteriorates due to lower frequency of fine matching
Solution Approach 1:
The system implements a feedback mechanism where odometry motion estimates continuously inform and refine the mapping process. The odometry thread provides real-time motion compensation that feeds into the mapping thread, allowing frequent updates without accuracy loss by constantly correcting for scanner motion based on accumulated observations.
Solution Approach 2:
The mapping system dynamically adjusts its operation by performing coarse motion compensation at high frequency through odometry while reserving fine matching and registration for lower frequency updates. This dynamic multi-rate approach maintains accuracy while enabling real-time performance.
3Manufacturing precision
If odometry motion estimation accuracy is improved to offset mapping accuracy penalties, then mapping accuracy is maintained, but latency increases due to tighter convergence conditions
Solution Approach 1:
The patent separates odometry and mapping into independent parallel threads with different convergence requirements. The odometry thread performs rapid motion estimation with relaxed convergence for low latency, while the mapping thread performs accurate registration with tighter convergence when needed, eliminating the need to compromise one for the other.
Solution Approach 2:
The system performs partial motion compensation through odometry at high frequency with relaxed accuracy requirements, then applies excessive fine matching through mapping at lower frequency. This partial-excessive approach maintains overall accuracy while minimizing latency by not requiring full accuracy at every time step.
Data Source
AI summary
Systems and method for updating an accumulated 3D map are disclosed. A current point cloud is received, which is an untwisted lidar point cloud captured over a current interval, each point in the current point cloud associated with feature data indicating a feature type of each point of the current point cloud. Respective subsets of the current point cloud are provided to a plurality of processing threads, wherein each processing thread operates on its point cloud subset in parallel with the other processing thread(s) to perform the following mapping operations: compare each point of its point cloud subset with features of the accumulated 3D map to identify a corresponding feature of the same feature type in the accumulated 3D map, compute a distance between each point of its subset and the corresponding feature in the accumulated point cloud, and determine a derivative of each distance with respect to lidar pose change. The derivatives computed by the plurality of processing threads are used to: (i) calculate a refined pose change estimate over the current time interval, and (ii) augment the accumulated 3D map with the untwisted current point cloud using the refined pose change estimate.