INFRA-3DRC Radar-Based Road-User Classification

Road-user classification using radar point-cloud data from the INFRA-3DRC automotive perception dataset.

Overview

This repository provides two approaches for classifying road users from annotated radar point-cloud data across the 25 INFRA-3DRC scenes:

• Random Forest — classification using statistical and geometric features extracted from radar point sets.
• PointNet++ — deep-learning classification directly from radar point sets.

Both implementations use a stratified 70% training, 15% validation, and 15% testing split.

Classes

The following INFRA-3DRC classes are considered:

• Adult
• Child
• Group
• Bicycle
• Motorcycle
• Car
• Bus
• Truck

Files

• randomForestClass_25scenes_HPC.py — Random Forest training and evaluation.
• pointNetPlusPlusClass_25scenes_HPC.py — PointNet++ training and evaluation.

Dataset

The INFRA-3DRC dataset is not included in this repository.

The original dataset is available from the official INFRA-3DRC dataset website:

https://fraunhoferivi.github.io/INFRA-3DRC-Dataset/

Evaluation

The implementations report:

• Accuracy
• Balanced accuracy
• Macro-F1
• Weighted-F1
• F1-score

Results

Results from the 25-scene Random Forest and PointNet++ experiments will be added here.

Author

Olivier Rukundo, Ph.D.
University of Limerick, Ireland

References

If you use this work, please also cite the original INFRA-3DRC dataset, Random Forest, and PointNet++ publications.

Results

Performance on the 25-scene INFRA-3DRC test partition:

Method	Accuracy	Balanced Accuracy	Macro-F1	Weighted-F1
Random Forest	97.22%	91.55%	93.19%	97.16%
PointNet++	70.59%	60.74%	57.82%	73.26%

Both methods were evaluated on the same number of samples (612).

Per-Class Performance

Test F1-score for each represented road-user class:

Road user	Random Forest	PointNet++
Adult	0.96	0.59
Bicycle	0.96	0.39
Bus	0.99	0.94
Car	0.99	0.84
Group	0.88	0.47
Motorcycle	0.82	0.24