Adaptive DBSCAN for LiDAR Point Cloud Clustering

A clean implementation of an adaptive-parameter DBSCAN pipeline designed for LiDAR-based perception systems in robotics and autonomous driving.

This project explores how density-based clustering can be made robust to LiDAR’s range-dependent point density, where traditional global-parameter DBSCAN often fails.

---

Motivation

Standard DBSCAN uses fixed global parameters:

• ε (neighbourhood radius)
• min_samples (minimum core density)

However, LiDAR point clouds are non-uniform:

• Point density decreases with distance
• Beam divergence affects spatial resolution
• 3D structure is often projected into 2D

A single global ε leads to:

• Fragmented far-field clusters
• Merged near-field objects
• Sensitivity to tuning

This project investigates adaptive ε and min_samples strategies that scale with local geometry or density statistics.

---

High-Level Pipeline

Typical LiDAR obstacle clustering flow:

1. Region-of-interest filtering (optional)
2. Voxel grid downsampling
3. Ground removal (e.g., RANSAC plane fitting)
4. Adaptive DBSCAN clustering
5. Post-processing (centroids, bounding boxes, filtering)

---

Repository Structure

adaptive-dbscan-core/
│
├── notebooks/        # Experimental notebooks and visualisations
├── scripts/          # Runnable pipeline entrypoints
├── src/              # Reusable implementation modules (in progress)
├── README.md
├── LICENSE
└── .gitignore

Large ROS bag files and raw LiDAR datasets are intentionally not tracked in Git.

Current Status

Implemented / explored:

• DBSCAN clustering in 2D projected LiDAR frames
• Adaptive epsilon scaling experiments
• Cluster visualisation with bounding boxes
• RANSAC ground segmentation experiments
• Voxel downsampling utilities
• (Data Extractor for converting RosBags into .npy files)

In progress:

• Extraction of clean core modules into src/
• Synthetic benchmarking
• Unit tests

Plans / Future Work

• World Model (Occupancy Grid): Maintain a 2D occupancy grid over the ground plane (or 3D voxel grid if needed). Each cell stores an occupancy probability representing belief that the region contains an obstacle.

• Bayesian Filter Updates (Log-Odds): Fuse LiDAR evidence across time using a recursive Bayesian update (log-odds form) with an inverse sensor model for "hit" and "free space" along each beam.

• Cluster on Aggregated Evidence: Threshold or sample occupied cells to produce a consolidated point set, then apply DBSCAN to recover consistent object clusters from the map rather than from a single frame.on that

---

Example Usage (Planned API)

from adaptive_dbscan.core import adaptive_dbscan

labels = adaptive_dbscan(
    points,                    # shape (N,2) or (N,3)
    eps_mode="adaptive",
    min_samples_mode="adaptive",
    device="cuda"
)