Evaluating feature extraction on wearable sensor data for outdoor sports activity recognition

Thumbnail Image

Files

master_Kassem_Donia_2026.pdf (1.3 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Electrical Engineering | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-12-28

Department

Major/Subject

Control, Robotics and Autonomous Systems

Mcode

Degree programme

Master's Programme in Automation and Electrical Engineering
Automaation ja sähkötekniikan maisteriohjelma
Magisterprogrammet i automation och elektroteknik

Language

en

Pages

65

Series

Abstract

Human Activity Recognition (HAR) using wearable sensors has become an essential component of modern health, fitness, and sports analytics systems. However, the traditional methods are usually based on the handcrafted features, which restrict generalization and adaptability in the dynamic-real world settings like outdoor sports. Optimized feature schemes have been popular in the domains of natural language processing and acoustic signal processing, however, their usage for the application of activity recognition has not been explored much. Therefore, this thesis evaluates the performance of standardized time-series feature extraction schemes—CATCH22, TSFEL, TSFRESH, and HCTSA—for outdoor sports activity recognition using inertial sensor data from the WEAR dataset. The dataset includes accelerometer recordings from multiple body placements (wrists and ankle) of 22 participants performing 18 outdoor activities across diverse terrains. After data preprocessing involving windowing, labeling, and relabeling into seven consolidated activity classes, features were extracted using the four schemes and refined through Recursive Feature Elimination (RFE) and Minimum Redundancy Maximum Relevance (MRMR) selection methods. The selected features were then classified using Support Vector Machines (SVM), Decision Trees (DT), Random Forest (RFC), and Extreme Gradient Boost (XGB) algorithms. The results demonstrate that the CATCH22 feature set, when combined with RFE and the XGB classifier, achieved the best overall classification performance with high accuracy and computational efficiency, especially when considering edge deployment. Comparatively, TSFRESH and HCTSA offered richer but more computationally demanding feature spaces. The findings confirm that optimized feature extraction libraries, particularly CATCH22, provide a balanced trade-off between interpretability, robustness, and real-time applicability for wearable HAR in outdoor sports. This work provides a systematic and practical evaluation framework that supports the selection of computationally efficient and deployable feature extraction strategies and contributes to advancing reproducible and efficient featurebased frameworks that improve recognition accuracy and adaptability in unconstrained, real-world athletic environments.

Description

Supervisor

Zhou, Quan

Keywords

human activity recognition, wearable sensors, time-series feature extraction, CATCH22, feature selection, outdoor sports

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202601191430

Collections

[dipl] Sähkötekniikan korkeakoulu / ELEC