Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks

Nengbo Zhang; Yao Ying; Lu Wang; Kaishun Wu; Jieming Ma; Fei Luo

cs.CVMar 15, 2026

Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks

Authors: Nengbo Zhang, Yao Ying, Lu Wang, Kaishun Wu, Jieming Ma, Fei Luo

Organizations: School of Computing and Information Technology, Great Bay University, Dongguan, China · School of Computer Science, Guangdong University of Petrochemical Technology, Maoming, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China · Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Department of Computer Science and Technology, Harbin Institute of Technology, Shenzhe, China

Abstract

WiFi-based human action recognition (HAR) has gained significant attention due to its non-intrusive and privacy-preserving nature. However, most existing WiFi sensing models predominantly focus on improving recognition accuracy, while issues of power consumption and energy efficiency remain insufficiently discussed. In this work, we present Wi-Spike, a bio-inspired spiking neural network (SNN) framework for efficient and accurate action recognition using WiFi channel state information (CSI) signals. Specifically, leveraging the event-driven and low-power characteristics of SNNs, Wi-Spike introduces spiking convolutional layers for spatio-temporal feature extraction and a novel temporal attention mechanism to enhance discriminative representation. The extracted features are subsequently encoded and classified through spiking fully connected layers and a voting layer. Comprehensive experiments on three benchmark datasets (NTU-Fi-HAR, NTU-Fi-HumanID, and UT-HAR) demonstrate that Wi-Spike achieves competitive accuracy in single-action recognition and superior performance in multi-action recognition tasks. As for energy consumption, Wi-Spike reduces the energy cost by at least half compared with other methods, while still achieving 95.83% recognition accuracy in human activity recognition. More importantly, Wi-Spike establishes a new state-of-the-art in WiFi-based multi-action HAR, offering a promising solution for real-time, energy-efficient edge sensing applications.

Figures & tables

Fig. 1 : Overview of WiFi sensing model based on Spiking Neural Network. Conventional WiFi-based human sensing systems, relying on artificial neural networks, suffer from high energy consumption. Inspired by brain-like models, we propose a novel WiFi human action recognition model by replacing traditional neurons with spiking neurons. This approach,termed Wi-Spike, achieves a low-power WiFi-based human action recognition system.

Fig. 2 : Description of the basic principles of Spiking Neural Networks and Artificial Neural Network.

Fig. 3 : A detailed comparison is provided in the single-action WiFi recognition and multi-action WiFi recognition problems. This includes WiFi CSI signals for one action, two actions, and three actions. In the right part of Fig. 3 , clearly, the more human motion patterns in the WiFi CSI, the greater the number of step signals.

Fig. 4 : Overview of Wi-Spike Architecture. From left to right, the raw WiFi CSI human activity signals sequentially pass through the spike encoder, spike feature extraction, spike feature encoding, and voting classification modules.

Fig. 5 : Detailed structure of the Temporal Attention Layer.

Panel of Figure 6 — Fig. 6 : Confusion matrices on NTU-Fi-HAR and NTU-Fi-HumanID datasets.

Algorithm	Accuracy	F1-score	Precision	Recall
ours	$96.97\%$	$96.93\%$	$97.06\%$	$96.90\%$
ViT	$93.18\%$	$93.02\%$	$93.95\%$	$93.08\%$
Resnet18	$95.19\%$	$95.20\%$	$95.21\%$	$95.19\%$
CNN+GRU	$92.05\%$	$91.92\%$	$93.19\%$	$92.05\%$
LSTM	$97.70\%$	$96.95\%$	$98.19\%$	$96.96\%$
CNN	$98.60\%$	$98.72\%$	$98.7\%$	$96.96\%$

TABLE I : The classification results on NTU-Fi-HAR dataset.

Algorithm	Accuracy	F1-score	Precision	Recall
ours	$95.58\%$	$95.45\%$	$96.14\%$	$95.58\%$
ViT	$78.91\%$	$78.38\%$	$79.34\%$	$78.91\%$
ResNet18	$96.63\%$	$96.45\%$	$97.26\%$	$96.54\%$
CNN+GRU	$87.27\%$	$86.40\%$	$88.37\%$	$87.07\%$
LSTM	$97.19\%$	$96.74\%$	$98.84\%$	$97.13\%$
CNN	$97.10\%$	$96.55\%$	$98.65\%$	$96.60\%$

TABLE II : The classification results on NTU-Fi-HumanID dataset.

Algorithm	Accuracy	F1-score	Precision	Recall
ours	$96.51\%$	$95.85\%$	$96.34\%$	$96.98\%$
ViT	$96.25\%$	$96.28\%$	$96.33\%$	$96.09\%$
ResNet18	$98.00\%$	$97.99\%$	$98.02\%$	$97.99\%$
CNN+GRU	$96.18\%$	$96.10\%$	$96.33\%$	$96.08\%$
LSTM	$87.45\%$	$86.62\%$	$88.73\%$	$87.14\%$
CNN	$97.61\%$	$97.65\%$	$97.15\%$	$96.97\%$

TABLE III : The classification results on UT-HAR dataset.

Fig. 7 : Comparison of Algorithm Accuracy across Different action lengths in NTU-Fi-HAR datasets.

Fig. 8 : Comparison of classification accuracy among different models on the UT-HAR datasets. The UT-HAR_one, UT-HAR_two, and UT-HAR_three datasets contain different sets of action information, allowing evaluation of model robustness under varying action configurations.

Algorithm	Accuracy	F1-score	Precision	Recall
ours	$95.83\%$	$95.79\%$	$96.38\%$	$95.83\%$
ViT	$86.54\%$	$86.40\%$	$86.60\%$	$86.65\%$
ResNet18	$89.69\%$	$86.99\%$	$90.44\%$	$87.50\%$
CNN+GRU	$84.38\%$	$76.38\%$	$88.12\%$	$81.06\%$
LSTM	$88.78\%$	$87.74\%$	$88.84\%$	$89.18\%$
CNN	$91.87\%$	$89.59\%$	$92.94\%$	$90.15\%$

TABLE IV : The classification experiment on NTU-Fi-HAR double dataset.

Algorithm	Accuracy	F1-score	Precision	Recall
ours	$92.12\%$	$91.02\%$	$93.58\%$	$92.12\%$
ViT	$70.29\%$	$70.78\%$	$71.34\%$	$70.41\%$
ResNet18	$90.75\%$	$88.06\%$	$92.82\%$	$89.46\%$
CNN+GRU	$77.50\%$	$76.88\%$	$69.05\%$	$78.21\%$
LSTM	$91.50\%$	$91.25\%$	$91.78\%$	$91.35\%$
CNN	$88.81\%$	$87.58\%$	$90.33\%$	$88.08\%$

TABLE V : The classification experiment on NTU-Fi-HumanID double dataset.

Fig. 9 : Firing rate of each network layer across training epochs for both training and testing phases.

Fig. 10 : Mean firing rate of each network layer during training and testing.

Fig. 11 : Comparison of model accuracy and energy consumption.

Algorithm	Input Size	Energy	Accuracy	MACs	ACs(M)	Param
ours	$3\times 114\times 500$	$0.07$	$95.83\%$	-	73.51 M	14.34 M
ViT	$3\times 114\times 500$	$2.33$	$86.40\%$	507.1 M	-	0.83 M
ResNet18	$3\times 114\times 500$	$0.25$	$89.69\%$	54.27 M	-	11.18 M
CNN+GRU	$3\times 114\times 500$	$0.24$	$84.38\%$	53.96 M	-	0.05 M
LSTM	$3\times 114\times 500$	$0.24$	$88.78\%$	52.54 M	-	0.10 M
CNN	$3\times 114\times 500$	$0.13$	$91.87\%$	28.28 M	-	0.47 M

TABLE VI : Detailed comparison of computational efficiency and power consumption in NTU-Fi-HAR_Double.

Fig. 12 : Comparison of accuracy and loss changes during neural networks training.

Panel of Figure 19 — Fig. 13 : Comparison of feature maps after first layer between Wi-spike (left) and traditional CNN (right)

Explore similar work

WISE-HAR: A Generalizable Ensemble Deep Learning Framework for WiFi-Based Human Activity Recognition
Jun 2, 2026Maheen Arshad, Qindeel E Zahra, Muhammad Khuram ShahzadHuman Activity Recognition Wi-Fi Sensing
Department of Computing, School of Electrical Engineering and Computer Science National University of Sciences and Technology (NUST), Islamabad, Pakistan
Physics-Guided Attention in a Lightweight TCN for Efficient WiFi CSI-Based Human Activity Recognition
Jun 1, 2026Chinthaka Ranasingha, Tharindu Fernando, Sridha Sridharan +2Human Activity Recognition Temporal Convolutional Networks
Signal Processing, Artificial Intelligence and Vision Technologies (SAIVT) Research Group, School of Electrical Engineering and Robotics, Queensland University of Technology (QUT), Brisbane, QLD 4000, Australia
WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features
Aug 1, 2026Alison M. Fernandes, Hermes I. Del Monego, Bruno S. Chang +3Human Activity Recognition Multimodal Fusion
Universidade Tecnol´ogica Federal do Paran´a (UTFPR), Programa de P´os-Graduac¸˜ao em Engenharia El´etrica e Inform´atica Industrial (CPGEI-CT), Av. Sete de Setembro, 3165, Curitiba, 80230-901, Brazil · INESC TEC, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, Porto, 4200-465, Porto, Portugal

Jun 2, 2026J/K move · Enter open · S save

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

Maheen Arshad, Qindeel E Zahra, Muhammad Khuram Shahzad

Department of Computing, School of Electrical Engineering and Computer Science National University of Sciences and Technology (NUST), Islamabad, Pakistan

Human Activity Recognition Wi-Fi Sensing

Open paper AskPDF arXiv

Human Activity Recognition (HAR) using WiFi signals has emerged as a transformative technology for smart homes, healthcare monitoring, security systems, and ambient assisted living. Unlike traditional camera-based systems that raise significant privacy concerns and fail in low-light conditions, or wearable sensors that require user compliance, WiFi-based HAR is non-intrusive, privacy-preserving, cost-effective, and works seamlessly in any lighting condition. This paper presents a comprehensive approach to recognize three distinct human activities: "No Presence" (empty room), "Walking", and "Walking + Arm-waving" using the Wallhack1.8k WiFi spectrogram dataset. We propose three key improvements to address the main challenges in WiFi-based HAR. First, to address high performance variance, we implement ensemble learning with five different CNN architectures (Deep CNN, Wide CNN, MobileNetV2, ResNet50V2, and EfficientNetB0). Second, to address the small dataset size limitation, we apply aggressive data augmentation techniques including time-warping, frequency masking, and noise addition. Third, to evaluate real-world generalization capability, we perform cross-scenario evaluation (training on Line-of-Sight and testing on Non-Line-of-Sight) and cross-antenna evaluation (training on Biquad antenna and testing on PIFA antenna). Our ensemble model achieved a test accuracy of 94.87% on the LOS scenario with Biquad antenna, outperforming the best individual model by 0.66%. Data augmentation improved Random Forest performance from 60% to 95%. Cross-scenario evaluation showed minimal accuracy drops of only 1.37% and 2.07%, demonstrating strong generalization capabilities. The results indicate that the proposed approach is robust, reliable, and suitable for real-world deployment in diverse environments with different hardware configurations.

Abstract

Explore similar work

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

Physics-Guided Attention in a Lightweight TCN for Efficient WiFi CSI-Based Human Activity Recognition

WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features