Quantitative Analysis of Mother Wavelet Function Selection for Wearable Sensors-Based Human Activity Recognition

Heba Nematallah ¹, Sreeraman Rajan ¹

⁰Department of Systems and Computer Engineering, Carleton University, Ottawa, ON K1S 5B6, Canada.

Sensors (basel, Switzerland) +

|

April 13, 2024

View abstract on PubMed

Summary

Selecting the right mother wavelet is crucial for accurate human activity recognition (HAR) using wearable sensors. This study proposes an optimal wavelet selection method to enhance HAR performance with minimal computational needs.

Area Of Science

* Signal Processing
* Machine Learning
* Wearable Technology

Background

* Advancements in Internet of Things (IoT) wearable devices, particularly inertial sensors, necessitate precise human activity recognition (HAR).
* Wavelet transform is suitable for HAR due to its time-frequency localization, but optimal mother wavelet selection is critical for performance.
* Activity signals exhibit distinct periodic patterns, making mother wavelet shape matching to sensor data vital for recognition accuracy.

Purpose Of The Study

* To propose and evaluate an optimal mother wavelet selection method for human activity recognition (HAR).
* To investigate the impact of mother wavelet selection on HAR performance using wavelet packet transform and entropy-based criteria.
* To provide guidelines for selecting optimal mother wavelets in HAR systems utilizing wearable inertial sensors.

Main Methods

* Implemented an optimal mother wavelet selection method combining wavelet packet transform with the energy-to-Shannon-entropy ratio.
* Utilized decision tree (DT) and support vector machines (SVM) as classification algorithms.
* Evaluated six mother wavelet families across eight public ADL datasets (MHEALTH, WISDM, HARTH, HARsense, DaLiAc, PAMAP2, REALDISP, HAR70+).

Main Results

* The proposed method demonstrated the significant impact of mother wavelet selection on HAR performance.
* Optimal wavelet selection, guided by signal shape resemblance and energy-entropy ratio, improved recognition accuracy.
* Analysis confirmed the effectiveness of the combined approach across diverse datasets.

Conclusions

* Optimal mother wavelet selection is a critical factor for enhancing HAR accuracy in wearable sensor systems.
* The energy-to-Shannon-entropy ratio effectively guides the selection of wavelets that match activity signal characteristics.
* This study offers a valuable framework for optimizing mother wavelet selection in HAR applications.

Keywords:

digital signal processing (DSP) human activity recognition (HAR) mother wavelet (MW) analysis wavelet packet transform (WPT) wavelet transform (WT)