Title: WISE-HAR: A Generalizable Ensemble Deep Learning Framework for WiFi-Based Human Activity Recognition

Abstract:

Human Activity Recognition (HAR) leveraging WiFi signals has evolved into a pivotal technology for applications ranging from smart homes and healthcare monitoring to security systems and ambient assisted living. In contrast to camera-based solutions, which are often hindered by privacy issues and poor performance in low-light environments, or wearable devices that demand strict user adherence, WiFi-based HAR offers a non-intrusive, cost-efficient, and privacy-friendly alternative that functions effectively regardless of lighting conditions.

This study introduces a robust methodology for identifying three specific activities—"No Presence" (indicating an empty room), "Walking," and "Walking + Arm-waving"—utilizing the Wallhack1.8k WiFi spectrogram dataset. To overcome the primary obstacles in WiFi-based HAR, we introduce three strategic enhancements. First, to mitigate high performance variance, we employ an ensemble learning strategy combining five distinct Convolutional Neural Network (CNN) architectures: Deep CNN, Wide CNN, MobileNetV2, ResNet50V2, and EfficientNetB0. Second, to counteract the limitations of small dataset sizes, we utilize aggressive data augmentation methods, such as time-warping, frequency masking, and noise injection. Third, to assess real-world generalization, we conduct cross-scenario evaluations (training on Line-of-Sight data and testing on Non-Line-of-Sight data) and cross-antenna evaluations (training on Biquad antenna data and testing on PIFA antenna data).

The proposed ensemble model attained a test accuracy of 94.87% in the Line-of-Sight scenario using the Biquad antenna, surpassing the highest-performing individual model by 0.66%. Additionally, data augmentation significantly boosted Random Forest performance, rising from 60% to 95%. The cross-scenario tests revealed minimal accuracy declines of just 1.37% and 2.07%, underscoring the model’s strong generalization potential. These findings suggest that the proposed approach is both robust and reliable, making it well-suited for real-world deployment across diverse environments and varying hardware configurations.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC