Pseudo-Image Instance Segmentation for Low-Compute Vehicle Perception

The patent divides the point cloud data into multiple pillars along the vertical dimension, transforming complex 3D data into structured 2D pillar representations. This segmentation enables parallel processing of different pillars independently, significantly improving processing throughput while reducing computational complexity per pillar through the use of efficient 2D convolutional neural networks

The patent transforms 3D point cloud data into 2D pillar images by projecting points onto a 2D grid with vertical height information. This dimensionality reduction from 3D to 2D while preserving essential spatial and height features enables the use of computationally efficient 2D CNN architectures, achieving real-time processing speeds with reduced computational resource consumption

The patent performs preliminary organization of point cloud data into structured pillars with predefined spatial relationships before processing. This pre-structuring establishes height maps, depth information, and spatial contexts in advance, enabling the neural network to focus computational resources on feature extraction and classification rather than basic data organization, thus reducing overall processing time while maintaining perception accuracy

The patent applies different processing strategies to different pillars based on their local characteristics. Each pillar is processed independently with localized feature extraction, allowing the system to adapt computation to local complexity. Regions with simpler structures require fewer computational resources while maintaining the same perception accuracy, thereby reducing total processing time across the entire scene