HIT-MHAD

Human activity recognition (HAR) plays a critical role in pervasive computing and human-centered applications. However, most existing datasets are constrained by single-sensor modalities and outdated Kinect versions with limited skeletal joints. In particular, there remains a gap incorporating flexible stretch sensors. To address these issues, we present HIT-Multimodal Human Action Dataset (HIT-MHAD), a multimodal human action dataset that includes 24-channel stretch sensor data synchronized with depth, infrared, and 3D skeleton data captured by an Azure Kinect device. These wearable stretch sensors offer natural movement compliance through elastic textile substrates and conform to human joint anatomy, enabling detailed full-body motion capture in vision-constrained environments. The dataset comprises 2,400 sequences with over 1,030,000 frames of full-body kinematic data and corresponding action labels. It covers 24 action classes, including gestures, daily activities, and fitness-related movements. We conduct extensive experiments to validate the dataset. Results indicate that HIT-MHAD supports both unimodal and multimodal evaluations effectively. It is suitable for several downstream tasks, such as 2D human pose estimation, multimodal sensor fusion for action recognition, pose estimation from wearables, and cross-modal learning. The dataset is publicly available at Zenodo: https:\/\/doi.org\/10.5281\/zenodo.16893501.