Augmented Point Cloud Generation via Temporal Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer perception systems using LiDAR and radar generate sparse 3D coordinate data, limiting their range and accuracy, especially at farther distances, and require recalculating point features at each time step, which is inefficient.
Innovation Solution
A method for generating a densified temporal point cloud by aggregating current and previous point clouds without quantization or voxelization, using a machine learning-based system to align and transform points, and incorporating neural networks for efficient reuse of previously computed features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sparse 3D coordinate data from DAR devices is used directly, then device complexity is reduced, but measurement precision deteriorates especially at farther distances
Solution Approach 1:
The system performs preliminary densification by aggregating current point cloud with historical point clouds before processing. This preliminary action of accumulating temporal data improves measurement precision at farther distances by creating a denser point cloud, while the aggregation operation itself adds manageable computational complexity rather than requiring complex hardware upgrades.
Solution Approach 2:
The invention transitions from spatial-only point cloud processing to spatio-temporal processing by incorporating the time dimension. Historical point clouds from multiple time steps are aggregated with the current point cloud, adding temporal accumulation as a new dimension that enhances measurement precision without fundamentally changing the DAR device hardware.
2Productivity
If point features are recalculated at each time step, then measurement precision is maintained, but productivity deteriorates due to redundant computation
Solution Approach 1:
Point features are calculated preliminarily at the first time step and then reused in subsequent time steps through the aggregation process. This preliminary calculation avoids redundant computation at each time step, significantly improving processing efficiency while maintaining feature accuracy through temporal aggregation of the already-computed features.
Solution Approach 2:
Instead of discarding and recalculating point features at each time step, the system recovers and reuses previously computed features by aggregating them with current data. This recovery approach eliminates redundant computation while preserving feature accuracy, directly addressing the productivity-precision trade-off.
3Measurement precision
If quantization or voxelization is applied to aggregate point clouds, then device complexity is reduced, but measurement precision deteriorates due to information loss
Solution Approach 1:
The invention extracts and removes the quantization and voxelization steps from the aggregation process. By taking out these information-lossy operations, the system directly aggregates raw 3D coordinate data points, preserving measurement precision and resolution while managing complexity through efficient point-wise aggregation algorithms.
Solution Approach 2:
Instead of applying quantization or voxelization to simplify aggregation (the conventional approach), the invention inverts the approach by aggregating raw high-precision data directly and only applying simplification if absolutely necessary. This inversion prioritizes measurement precision while handling complexity through smarter aggregation logic.
Data Source
AI summary
A method for generation of an augmented point cloud with point features from aggregated 3D coordinate data and related device. The method comprises receiving a current point cloud in the form of 3D coordinate data in ego coordinates from one or more detection and ranging (DAR) devices of a vehicle. Features are extracted from the current point cloud. A previous point cloud is transformed into ego coordinates using a current location of the vehicle. Each point in the previous point cloud is transformed to align with a corresponding point in the current point cloud to generate a transformed point cloud. The current point cloud is aggregated with the transformed point cloud to generate an aggregated point cloud. The current point features are aggregated with the point features of the transformed point cloud to generate aggregated point features.