Hybrid Deep Learning Framework for CSI-Based Activity Recognition in Bandwidth-Constrained Wi-Fi Sensing

TL;DR

This work addresses the challenge of reliable CSI-based human activity recognition under bandwidth constraints. It introduces IBIS, a hybrid framework that first enhances motion-sensitive features via Doppler trace extraction and then processes the results with a spatial Inception module, a temporal BiLSTM, and a final SVM classifier, with Grid Search tuning of kernel and hyperparameters. Experimental results across 20, 40, and 80 MHz demonstrate strong robustness, achieving 89.27% accuracy at 20 MHz and up to 95.30% at 80 MHz, outperforming a baseline Inception model, especially in low-bandwidth scenarios. The study highlights the value of combining Doppler-based feature engineering with hybrid learning and antenna diversity for practical, bandwidth-constrained Wi-Fi sensing HAR applications.

Abstract

This paper presents a novel hybrid deep learning framework designed to enhance the robustness of CSI-based Human Activity Recognition (HAR) within bandwidth-constrained Wi-Fi sensing environments. The core of our proposed methodology is a preliminary Doppler trace extraction stage, implemented to amplify salient motion-related signal features before classification. Subsequently, these enhanced inputs are processed by a hybrid neural architecture, which integrates Inception networks responsible for hierarchical spatial feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) networks that capture temporal dependencies. A Support Vector Machine (SVM) is then utilized as the final classification layer to optimize decision boundaries. The framework's efficacy was systematically validated using a public dataset across 20, 40, and 80 MHz bandwidth configurations. The model yielded accuracies of 89.27% (20 MHz), 94.13% (40 MHz), and 95.30% (80 MHz), respectively. These results confirm a marked superiority over standalone deep learning baselines, especially in the most constrained low-bandwidth scenarios. This study underscores the utility of combining Doppler-based feature engineering with a hybrid learning architecture for reliable HAR in bandwidth-limited wireless sensing applications.

Figures (6)

• Figure 1: IBIS processing pipeline: CSI signals are collected and transformed via Doppler trace extraction, followed by spatial–temporal feature learning through Inception–BiLSTM networks and final activity classification using an SVM.
• Figure 2: CSI acquisition setup: wireless signals are transmitted and received via multiple antennas, enabling CSI extraction for subsequent Doppler-based processing and activity recognition.
• Figure 3: Normalized Doppler spectrograms for five activities—Empty, Sitting, Walking, Running, and Jumping—captured at three bandwidth configurations: (a) 20 MHz, (b) 40 MHz, and (c) 80 MHz. Higher motion levels introduce greater Doppler variation, increasing classification complexity.
• Figure 4: Classification accuracy of Inception and IBIS across bandwidths of 20, 40, and 80 MHz. IBIS consistently outperforms the baseline, with the largest gain at 20 MHz.
• Figure 5: Average classification accuracy of Inception and IBIS as a function of the number of antennas. IBIS benefits significantly from antenna diversity, achieving its highest performance with four antennas.