DynFilter3D

Dynamic object filtering for 3D LiDAR point clouds using spatiotemporal geometry

Status: Work in progress. The API may change without notice.
TODO: Add a teaser image or short demo video illustrating the algorithm.

Introduction

dynfilter3d is a clean, from-scratch Python/C++ toolkit for identifying and filtering dynamic points in LiDAR range data using spatiotemporal information such as surface normals and short-term motion cues.

The design of this project is inspired by the method described in:

Falque, R., Le Gentil, C., & Sukkar, F.
Dynamic Object Detection in Range Data using Spatiotemporal Normals.
Australasian Conference on Robotics and Automation (ACRA), 2023.
https://github.com/UTS-RI/dynamic_object_detection

Motivation

• Provide a simple, easy-to-use library/framework for dynamic filtering.
• Separate logic from ROS dependency.
• Add utilities for visualization, benchmarking, and dataset testing.

Features

• Lightweight C++17 core for fast geometric processing.
• Python bindings via nanobind.
• Dynamic point detection using spatiotemporal normals.
• High-level Python API for both offline and streaming use.
• Optional visualization utilities (via open3d).
• Examples on open datasets (e.g. KITTI).

---

Quick Start (example usage)

import numpy as np
import dynfilter3d as df

# Load a small sequence of LiDAR scans
scans = [
    np.load("scan_000.npy"),
    np.load("scan_001.npy"),
    np.load("scan_002.npy"),
]

# Create detector
f = df.DynamicFilter3D(window_size=3, min_neighbors=8)

# Estimate temporal surface statistics
f.fit_sequence(scans)

# Filter dynamic points in the last scan
mask = f.get_static_points(idx=-1)
static_points = scans[-1][mask]

print("Static points:", static_points.shape[0])

The API is still evolving, but this illustrates the intended workflow.

Installation (from source)

git clone https://github.com/<your-username>/dynfilter3d.git
cd dynfilter3d

python -m venv .venv
source .venv/bin/activate

pip install -U pip
pip install -e .

This builds the native C++ extension and installs the Python package.

Command Line Interface (planned)

dynfilter3d \
    --input "sequence/*.pcd" \
    --output "filtered/" \
    --config config/default.yaml

Datasets

The package supports LiDAR point clouds in .pcd, .bin, .npy, and custom formats. Example scripts will be provided for:

• KITTI odometry,
• open-source urban LiDAR datasets,
• synthetic demo point clouds.

---

License

MIT License — a permissive, open-source license allowing commercial and academic use.

Acknowledgement

This project is inspired by the original method:

@inproceedings{falque2023dynamic,
    title={Dynamic Object Detection in Range data using Spatiotemporal Normals}, 
    author={Raphael Falque and Cedric Le Gentil and Fouad Sukkar},
    booktitle={Australasian Conference on Robotics and Automation, ACRA},
    year={2023}
}