Machine Learning-Based Real-Time Walking Activity and Posture Estimation in Construction With a Single Wearable Inertial Measurement Unit

Construction workers regularly perform walking locomotion on level and inclined surfaces. It is critical to detect walking activity and estimate body postures in real time for monitoring workers’ safety and health conditions. This article presents a machine learning-based framework for real-time activity detection and posture estimation during human walking on level and sloped terrains using a single wearable inertial measurement unit (IMU). The framework integrates recurrent neural networks with Gaussian process dynamical models to achieve accurate predictions of walking activity, floor slope angles, and workers’ turning angles and full-body limb joint angles estimation in real time. The proposed design offers a streamlined, cost-effective solution with significant advantages over multi-sensor systems. Extensive experiments of different walking activities on level and sloped surfaces are conducted to validate and demonstrate the design. The proposed algorithm detects gait activities with 96% accuracy, the estimated human limb joint angle errors are within 11 deg, the predicted turning angles have an error less than 16 deg and the end-to-end detection latency is within 21 ms using only one single IMU attached to the human shank. Note to Practitioners—Construction workers exert intense physical effort, and they experience serious safety and health risk in hazardous, dynamic working environments. As a result, the construction industry is one of the highest-risk industrial sectors in most countries. Real-time monitoring of workers’ walking gait plays a critical role in construction safety. Wearable IMUs are particularly attractive for walking activity recognition because they are small, inexpensive, and non-intrusive. This study aims to develop machine learning-enabled, real-time IMU-based walking activity recognition and full-body posture and floor slope angle estimation. The main approach is built on long short-term memory (LSTM)-based recurrent neural networks and a manifold learning method to process time-series data from the IMU to predict human motion. The LSTM model achieves over 96% accuracy to classify various walking activities on different floor slopes. Additionally, the design also estimates human limb joint angles, floor slope angle and the worker’s body turning angle. A noteworthy aspect of the detection system is the minimal detection latency of 18 ms, ensuring the reliability and effectiveness of real-time monitoring and evaluation. This feature is particularly beneficial for immediate feedback and intervention systems that help protect workers from work-related injuries. The simplicity and efficiency of using a single IMU are attractive for a practical solution for real-time automation applications in dynamic environments.