A Deep Learning-based Approach for Human Posture Classification

Abstract

Lifting posture is considered as a leading factor in low back injuries in the workplace. Hence, it is necessary to evaluate the risk of various lifting tasks. Classifying postures is important before performing an ergonomic task assessment. Recently, many studies have revealed that the deep learning method has a high accuracy in identifying human postures. However, few studies have explored how the deep learning method can be applied to classify different postures during a lifting task. The objective of this study was to develop a deep learning technique-based model for classifying three states of postures (squatting, standing and stooping) during a lifting task. A dataset comprising 2,600 various static images (squatting, standing and stooping) taken from 0° and 90° camera view angles and their corresponding 3D joint coordinate data recorded by the marker-based motion tracking system was used in this study. The images were randomly divided into training (1,300 images), validation (650 images) and testing (650 images) datasets. After all of the images were cropped to a fixed size, the training dataset was processed in the neural network as the input, and the validation dataset was used to revise the weight of the model while training to build the classifying model. Finally, the testing dataset was processed as input for classifying three static postures using the proposed model. A classification based on the 3D coordinate data captured by the marker-based motion tracking system was used as the reference to validate the accuracy of this classifying model. Overall, the model developed in this study reached 91.23% accuracy. The accuracy of correctly classifying the squatting, standing and stooping postures is 94.35%, 98.33% and 75.86%, respectively. In addition, this model showed a nearly equivalent accuracy for identifying the images taken from 0° (91.64%) and 90° (90.86%) cameras. The results of this preliminary test showed that the deep learning method has the potential to classify different static postures within a lifting pattern.