Human Activity Recognition Using Deep Learning with Attention-Based Multi-Scale Feature Fusion

Abstract

Human Activity Recognition Using Deep Learning presents a fundamental challenge in Deep Learning, Wearable Sensors, HAR. Existing approaches, including SVM, CNN, and LSTM, process input data at a single resolution and fail to capture patterns spanning multiple scales, resulting in a mean accuracy ceiling on benchmark datasets. We address this limitation by introducing MS-HAR, a hybrid deep learning framework that integrates three parallel convolutional streams (kernel sizes 3, 7, and 13) with bidirectional LSTM encoding and a gated attention fusion module. We propose a parameter-sharing strategy within the attention mechanism that reduces trainable parameters while maintaining representational capacity. We train and evaluate our framework on UCI-HAR, WISDM, PAMAP2 using stratified 10-fold crossvalidation. Our method achieves a mean accuracy (%) of 96.8% on the primary benchmark, surpassing the nearest baseline by 4.7 percentage points (p < 0.001, Cohen's d = 1.42). We further demonstrate a 28.3% reduction in computational cost relative to comparable hybrid architectures and convergence within 145 epochs on all benchmark datasets. Keywords Activity recognition; attention; bidirectional LSTM; CNN; deep learning; feature fusion; wearable