2.5D Adaptive LiDAR Processing Pipeline
A sensor-centric, distance-adaptive perception pipeline that transforms dense 3D LiDAR point clouds into a compact, semantically rich 2.5D representation while preserving detail where it matters.
RandLA-Net
A point-cloud segmentation network used to classify the environment at point level. Its semantic output provides the foundation for separating terrain, static structures, and dynamic objects before rasterization.
DBSCAN
A density-based clustering method applied to segmented spatial regions to separate individual object instances. It does not require a predefined number of clusters and can identify irregularly shaped groups while rejecting sparse noise.
Point-Cloud Preprocessing
Prepares the incoming LiDAR point cloud for downstream perception by reducing unnecessary point-level redundancy while preserving essential scene geometry.
- Spatial alignment & normalization
- Noise & outlier removal
- Point-cloud preparation
Semantic Segmentation
Classifies the 3D environment point-by-point to establish semantic boundaries for downstream layer separation.
- Drivable vs. non-drivable terrain
- Static structures & barriers
- Dynamic objects
Object Detection
Groups segmented points into spatial clusters to identify individual obstacle instances without requiring a predefined number of objects.
- Object spatial extent
- Height boundaries
- Semantic class association
Sensor-Centric Radial Grid
Divides the environment into concentric radial bands whose spatial resolution changes according to distance from the sensor.
Adaptive 2.5D Rasterization
Maps segmented and clustered scene information onto the radial grid, condensing point-level geometry into multi-layer summaries for each occupied cell.
2.5D Data Structure
Organizes the rasterized environment as a sparse radial structure. Empty cells are omitted rather than represented explicitly.
- Occupied cell IDs
- Layer-presence information
- Semantic metadata
Compact Binary Encoding
Packs the multi-layer cell representation into a compact fixed-size binary payload for efficient storage and transmission.
- Ground elevation & variation
- Static & dynamic height ranges
- Layer presence & semantic metadata
Compression & Streaming
Compresses the compact sparse representation for efficient storage and data transmission.
- Zstandard compression
- Sparse binary stream
- Incremental frame transfer
Visualization
Decodes the compact representation and renders the resulting radial 2.5D environment for interactive inspection.
- 3D displaced terrain & obstacles
- 2D semantic top-down mapping
- Distance-adaptive radial view