Human activity recognition from sensor data using spatial attention-aided CNN with genetic algorithm.

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Human Activity Recognition (HAR) from wearable sensors faces challenges in capturing temporal and frequency signal relationships.
• Current models struggle with spatiotemporal context extraction and large feature maps, impacting classification accuracy.

Purpose of the Study:

• To develop a novel hybrid architecture for improved HAR from wearable sensor data.
• To address limitations in spatiotemporal feature extraction and classification accuracy.

Main Methods:

• Utilized Continuous Wavelet Transform to convert time-series sensor data into multi-channel images.
• Employed a Spatial Attention-aided Convolutional Neural Network (CNN) for high-dimensional feature extraction.
• Developed a novel feature selection (FS) method combining filter-based techniques (MI, Relief-F, mRMR) and a modified Genetic Algorithm (GA).
• Classified activities using the K-Nearest Neighbors (KNN) classifier.

Main Results:

• The proposed hybrid model significantly outperformed state-of-the-art models on five benchmark HAR datasets (UCI-HAR, WISDM, MHEALTH, PAMAP2, HHAR).
• The GA-based FS technique improved overall recognition accuracy while reducing the number of features.
• Demonstrated superior classification performance compared to existing methods.

Conclusions:

• The hybrid architecture effectively captures complex time-frequency relationships for HAR.
• The integration of Continuous Wavelet Transform, Spatial Attention CNN, and GA-based FS offers a robust solution for wearable sensor-based HAR.
• The model provides a significant advancement in HAR accuracy and efficiency.