[Paper Review] B-HAR: an open-source baseline framework for in depth study of human activity recognition datasets and workflows

Florenc Demrozi, Cristian Turetta+1 more|arXiv (Cornell University)|Jan 23, 2021

Context-Aware Activity Recognition Systems9 citations

TL;DR

B-HAR is an open-source, configurable framework that provides a standardized, reproducible workflow for Human Activity Recognition (HAR) research, integrating common data preprocessing steps and machine/deep learning models. It enables researchers to systematically evaluate and compare HAR methodologies across diverse datasets while minimizing errors from incorrect processing pipelines, with validation on nine benchmark HAR datasets using inertial, physiological, and environmental sensors.

ABSTRACT

Human Activity Recognition (HAR), based on machine and deep learning algorithms is considered one of the most promising technologies to monitor professional and daily life activities for different categories of people (e.g., athletes, elderly, kids, employers) in order to provide a variety of services related, for example to well-being, empowering of technical performances, prevention of risky situation, and educational purposes. However, the analysis of the effectiveness and the efficiency of HAR methodologies suffers from the lack of a standard workflow, which might represent the baseline for the estimation of the quality of the developed pattern recognition models. This makes the comparison among different approaches a challenging task. In addition, researchers can make mistakes that, when not detected, definitely affect the achieved results. To mitigate such issues, this paper proposes an open-source automatic and highly configurable framework, named B-HAR, for the definition, standardization, and development of a baseline framework in order to evaluate and compare HAR methodologies. It implements the most popular data processing methods for data preparation and the most commonly used machine and deep learning pattern recognition models.

Motivation & Objective

• To address the lack of a standardized, reproducible workflow for evaluating HAR methodologies, which hinders fair comparison across studies.
• To reduce common errors in HAR pipeline design—such as incorrect processing order or misconfigured hyperparameters—by providing a configurable baseline framework.
• To support both off-the-shelf and user-defined machine and deep learning models while preserving consistent preprocessing across experiments.
• To improve transparency and reproducibility in HAR research by enforcing detailed documentation of data sources, processing steps, and model configurations.
• To enable researchers to fairly compare model performance across datasets by standardizing data preparation and evaluation procedures.

Proposed method

• B-HAR implements a modular, configurable pipeline that orchestrates data preprocessing steps including noise filtering, segmentation, feature extraction, normalization, feature selection, and class balancing.
• The framework supports integration of multiple machine learning and deep learning models, including CNNs, RNNs, and traditional classifiers, with configurable hyperparameters.
• It allows users to plug in custom models while reusing predefined preprocessing workflows, ensuring consistency and reducing implementation errors.
• The framework is designed with extensibility in mind, enabling future integration of new data processing techniques, models, and performance metrics.
• B-HAR enforces best practices by requiring explicit configuration of key parameters such as window size, overlap, normalization method, and train/test split strategy.
• It is validated across nine widely used HAR datasets, covering diverse sensor types (inertial, physiological, environmental) and activity categories.

Experimental results

Research questions

• RQ1 How can a standardized, reproducible workflow improve the comparability and reliability of HAR model evaluations across different datasets and research groups?
• RQ2 What are the critical data processing steps that commonly lead to errors or performance degradation in HAR pipelines, and how can they be systematically managed?
• RQ3 To what extent can a configurable open-source framework reduce implementation errors and improve reproducibility in HAR research?
• RQ4 How does the integration of user-defined models within a standardized preprocessing pipeline affect model performance and experimental fairness?
• RQ5 What level of configurability is needed to support both off-the-shelf models and novel architectures while maintaining consistency and transparency?

Key findings

• B-HAR successfully supports the implementation and evaluation of HAR workflows across nine benchmark datasets, demonstrating broad compatibility with diverse sensor modalities and activity types.
• The framework enables consistent and reproducible application of preprocessing steps such as segmentation, normalization, and feature selection, reducing variability due to manual configuration errors.
• By enforcing structured configuration of hyperparameters and processing steps, B-HAR minimizes common pitfalls such as incorrect train/test splits or improper data balancing.
• The integration of user-defined models within the framework’s standardized pipeline allows for fair comparison of novel models against established baselines without re-implementing preprocessing.
• Validation across multiple datasets confirms that B-HAR maintains high reproducibility and configurability, supporting both standard and advanced HAR research workflows.
• The framework provides a foundation for future extensions, including GUI support, new performance metrics, and integration of transfer learning and regression models.