Real-Time Human Activity Recognition on Edge Devices for Industrial Applications

Human Activity Recognition (HAR) is increasingly used to enhance exoskeleton control, collaborative robotics, and ergonomic assessments by providing high-level motion insights for industrial applications. However, deploying HAR on portable edge devices with limited computational resources presents challenges, highlighting the need for lightweight algorithms tailored to industry. This study proposes and compares low-powered lightweight HAR models capable of real-time operation on portable wearable systems. The models exploit sensor fusion from a full-body wearable system, the Smart Suit, designed for construction workers, with the goal of enabling accurate real-time activity recognition while preserving the Natural Human Motion Frequency (NHMF) of 5Hz. The impact of window size selection and the inclusion of transitional activities was analyzed to evaluate multiple Machine Learning (ML) (SVM, NN, Ensemble) and Deep Learning (DL) (CRNN, BiCRNN, BiLSTM, CNNTransformer) models combined with signal optimization techniques. Performance was assessed in terms of accuracy, F1-score, training time, inference time, FLOPs, energy consumption, and model size. A dedicated industrial dataset was iteratively constructed using a novel Elementary–Transition–Elementary (ETE) recording technique, which simplified the recording and labeling of complex tasks and supports dataset scalability. Model generalization was evaluated using a Leave-One-Subject-Out protocol, demonstrating that selected models achieved a maximum accuracy of 92.7% under realistic deployment constraints, enabling up to 10 classifications per second.