9th ABC Challenge - Dataset and Tutorial

Dataset Information

This challenge uses Ultra-Wideband (UWB) radar recordings collected at the LIARA Laboratory, Université du Québec à Chicoutimi (UQAC).

The recordings were collected in an approximately 40 m² instrumented apartment containing a kitchen, living room, bedroom, and bathroom.

The source dataset combines two complementary corpora:

1. Fall Corpus, containing recorded fall events;

2. Activities of Daily Living (ADL) Corpus, containing normal everyday activities such as walking, drinking, sleeping, cleaning, eating, reading, and using a computer.

The sensing setup consists of three wall-mounted Novelda XeThru X4M200 UWB radars. Each CIR frame contains 368 signal values, together with a timestamp and radar identifier. The effective sampling frequency is approximately 50 Hz per radar.

Challenge Package

The Data package is approximately 8.2 GB and is provided as a ZIP archive under the CC BY-NC 4.0 license.

It contains:

- 13 continuous training streams (~6.2 hours);

- 5 validation streams (~2 hours);

- Activities of Daily Living and fall events;

- temporal ground truth for 478 annotated falls;

- README with task, rules, data format, and evaluation protocol;

- official evaluation script;

- official submission format;

- two baseline methods;

- starter notebook.

The challenge uses continuous, non-segmented radar streams, requiring participants to locate falls and estimate their temporal boundaries rather than receiving pre-segmented fall intervals.

Dataset Structure

The source dataset consists of two main corpora:

dataset/

│

├── falls/

│ ├── Dataset 1

│ ├── Dataset 2

│ └── Dataset 3

│

└── activities/

├── Boire

├── Dormir

├── Enfiler_une_veste

├── Faire_menage

├── Faire_pates

├── Faire_the

├── Faire_vaisselle

├── Laver_dents

├── Laver_mains

├── Lire_livre

├── Manger

├── Marcher

├── Mettre_des_chaussures

├── Prendre_medicaments

└── Utiliser_ordinateur

Structure Description

The Fall Corpus contains 631 JSONL files, including 601 fall sequences, 20 other recordings, and 10 empty recordings. The fall events are divided into four fall types: Chute01, Chute02, Chute03, and Chute04.

The ADL Corpus contains 150 recordings, representing 15 activities performed by 10 subjects.

For challenge, these source recordings are used to construct continuous streams containing both ADLs and fall events, so participants must detect falls without receiving pre-segmented fall intervals.

Published papers about the dataset

J. Maitre, K. Bouchard, and S. Gaboury, “Fall Detection With UWB Radars and CNN-LSTM Architecture." IEEE Journal of Biomedical and Health Informatics, Vol. 25, No. 4, pp. 1273–1283, 2021. DOI: 10.1109/JBHI.2020.3027967
T. Imbeault-Nepton, J. Maître, K. Bouchard, and S. Gaboury, “Fall Detection from UWB Radars: A Comparative Analysis of Deep Learning and Classical Machine Learning Techniques.” ACM GoodIT 2023, Lisbon, Portugal, 2023., DOI: 10.1145/3582515.3609535
S. Abudalfa and K. Bouchard, “Transfer Learning using Computer Vision Models for Fall Detection from UWB Radars." Transportation Research Procedia, Vol. 84, pp. 19–26, 2025. DOI: 10.1016/j.trpro.2025.03.039
S. Abudalfa and K. Bouchard, “A Hybrid Vision Transformer and Residual Neural Network Model for Fall Detection Using UWB Radars." Applied Intelligence, Vol. 55, Article 222, 2025. DOI: 10.1007/s10489-024-06156-9