Temporal and Modality Awareness-Based Lightweight Residual Network with Attention Mechanism for Human Activity Recognition Using a Lower-Limb Exoskeleton Robot

Abstract

Although many human activity recognition (HAR) models have achieved high accuracy, their computational complexity often limits deployment in systems with constrained hardware resources, such as wearable lower-limb exoskeletons. In addition, existing models frequently overlook the complementary nature of multimodal sensor signals, focusing primarily on temporal dynamics while underutilizing modality-specific information. To address these issues, this study proposes a lightweight residual network for recognizing diverse locomotion activities across varying terrains using multimodal sensing data. The model adopts an asymmetric convolutional architecture composed of depthwise and pointwise layers to efficiently capture temporal and modality-specific features while significantly reducing the number of trainable parameters. A channel-attention block is further integrated to emphasize salient fused features. Evaluations on the Walking Assist Wearable Robot Motion dataset, which contains kinematic and postural signals from 500 adults using a lower-limb exoskeleton, demonstrated that the proposed model achieves an accuracy of 98.23% and a macro F1 score of 98.21%, with only 48,037 parameters. This outperforms four hybrid deep learning baselines while reducing the parameter count by 5.5-12.9 times. To assess generalizability, additional experiments were conducted on four benchmark datasets—UCI-HAR, HAPT, PAMAP2, and WISDM—under varying batch size conditions. The proposed model consistently achieved competitive or superior macro F1 scores of 0.9627±0.0046, 0.776±0.0138, 0.8937±0.0121, and 0.9736±0.0025, respectively, confirming its robustness and adaptability across diverse real-world HAR scenarios.