Deep Learning and Intelligent Perception

CLC number: TP391.4

Human Activity Recognition Using a CNN with an Enhanced Convolutional Block Attention Module

基于增强型CBAM 与卷积神经网络的人体活动识别方法

¹ School of Computer and Artificial Intelligence, Hefei Normal University, Hefei 236032, Anhui, China (合肥师范学院　计算机与人工智能学院，安徽　合肥　236032)
² Anhui Engineering Laboratory for Sports Health Information Monitoring Technology (AEL－SHIMT), Hefei 236032, Anhui, China (安徽省运动健康信息监测技术工程研究中心（合肥师范学院），安徽　合肥　236032)

Received: 25 July 2025

Abstract

WiFi-based human activity recognition (HAR) provides a non-intrusive approach for ubiquitous monitoring; however, achieving both high accuracy and robustness simultaneously remains a significant challenge. This paper proposes a Convolutional Neural Network with Enhanced Convolutional Block Attention Module (CNN-ECBAM) framework. The approach systematically converts raw Channel State Information (CSI) into pseudo-color images, effectively preserving essential signal characteristics for deep neural network processing. The core innovation is an Enhanced Convolutional Block Attention Module (ECBAM), tailored to CSI data characteristics, which integrates Efficient Channel Attention (ECA) and Multi-Scale Spatial Attention (MSSA). By employing learnable adaptive fusion weights, it achieves dynamic synergy between channel and spatial features, enabling the network to capture highly discriminative spatiotemporal patterns. The ECBAM module is integrated into a unified Convolutional Neural Network (CNN) to form the overall CNN-ECBAM model. Experimental results on the UT-HAR and NTU-Fi_HAR datasets demonstrate that CNN-ECBAM achieves competitive performance in recognition accuracy and outperforms mainstream baseline models. Specifically, it attains 99.20% accuracy on UT-HAR (surpassing ResNet-18 at 98.60%) and achieves 100% accuracy on NTU-Fi_HAR (exceeding GAF-CNN at 99.62%). These results validate the effectiveness of the proposed method for high-precision and reliable WiFi-based HAR.

摘要

基于 WiFi 的人体活动识别为普适监测提供了一种非侵入式方法，然而同时实现高精度和高鲁棒性的监测仍然是一个重大挑战。本文提出一种基于增强型卷积块注意力机制的卷积神经网络框架(CNN-ECBAM)，将原始信道状态信息系统性地转换为伪彩色图像，有效保留了深度神经网络处理所需的重要信号特征。核心创新在于设计了一种针对 CSI 数据特性的增强型卷积块注意力模块(ECBAM)，该模块融合了高效通道注意力(ECA)与多尺度空间注意力(MSA)，并通过可学习的自适应融合权重实现两者的动态协同，使网络能够捕获更具区分性的空间与通道特征。ECBAM模块被集成到统一的卷积神经网络(CNN)中，形成整体的 CNN-ECBAM 模型。在UT-HAR和NTU-Fi_HAR数据集上的实验结果表明，CNN-ECBAM 在识别准确率上取得了具有竞争力的性能，并超越了主流基准模型。在UT-HAR上准确率达99.2%(高于ResNet-18的98.6%)，在NTU-Fi_HAR上准确率达100%(高于GAF-CNN的99.62%)。这些结果验证了该方法在高精度、高可靠WiFi-HAR中的有效性。