Skip to content
Back to projects

Technical project

Point Cloud Classification

Geometric pipeline to classify LiDAR data into ground, buildings, and vegetation.

  • Python
  • LiDAR
  • PCA
  • KD-Tree

Overview

Built a full geometric workflow to classify a raw LAZ point cloud, generate a terrain model, and compare the output against AHN reference data.

Context

Developed as part of GEO1015 Digital Terrain Modelling at TU Delft, this project focused on building an interpretable classification pipeline using geometric methods rather than black-box models.

Objective

To classify a raw point cloud into ground, buildings, and vegetation, generate a DTM from classified ground points, and evaluate the output against AHN reference data.

Method

  1. 01

    Ground extraction using the Cloth Simulation Filter (CSF) algorithm

  2. 02

    DTM generation via Laplace interpolation on ground-classified points

  3. 03

    Local geometric feature extraction using PCA on KD-tree–derived neighborhoods

  4. 04

    Building detection via planar seed extraction, normal-constrained region growing, and RANSAC-based plane fitting with smoothing

  5. 05

    Vegetation detection using PCA-derived sphericity features, region growing on scattered neighborhoods and return attributes, and consensus-based refinement with smoothing

Visuals

Geometric classification strategy combining PCA-based local features with region growing and sampling consensus.
Building extraction using planar segmentation and region growing with normal constraints.
Vegetation classification using sphericity and neighborhood scattering features.
Comparison with AHN reference classification showing strong agreement and overclassification zones.

Analysis

  • Planarity and normal-angle thresholds strongly affected building extraction
  • Sphericity alone overgrew into buildings, so return information became important for vegetation
  • Performance was strongest in open terrain and weaker around rivers and data voids

Results

Normalized confusion matrix comparing predicted classification with AHN reference data. Ground accuracy reaches 97.7% and non-ground 93.1%, with minor confusion at class boundaries and complex surfaces.

Reflection

  • The project showed that geometric methods can perform strongly without deep learning
  • Runtime and memory constraints required chunk-based processing
  • Parameter tuning had a major influence on classification quality
  • The workflow also improved collaboration and GitHub-based project organization