Sports Activity Recognition Using Textile‐Based Sensor Fusion and Complex‐Valued Residual Convolutional Neural Networks

Traditional human activity recognition systems in wearable technologies typically separate sensing modalities, using mechanical sensors for motion detection and biophysical sensors such as electrocardiography for health monitoring. This work challenges that division by demonstrating that electrocardiographic signals captured via textile electrodes inherently encode activity‐specific information, as motion‐induced signal deformations, traditionally dismissed as noise, carry discriminative patterns. A smart textile system was developed integrating textile‐based electrocardiographic electrodes and capacitive strain sensors for multimodal recognition of 12 sport‐related activities. The framework employs a hierarchical sensor fusion strategy: Conventional machine learning models analyze strain data for postural recognition, while residual complex‐valued convolutional neural networks process electrocardiographic signals, marking their novel application in this context. Operating on wavelet‐transformed complex signals, the network extracts both amplitude and phase information, and incorporates component‐wise max pooling to improve computational efficiency without compromising accuracy. Experimental validation was conducted under intrasubject (three subjects) and intersubject (seven subjects) conditions, showing strong unimodal results. Strain‐based classifiers achieved 96.73% intra and 89.62% inter accuracy. Fusion meta‐classifiers with the contribution of the residual complex‐valued network yielded the highest performance, reaching 97.82% intra and 95.54% inter accuracy, demonstrating robust generalization and confirming the value of motion‐induced electrocardiographic features for activity recognition.