This study performed a clustering analysis using a human pose clustering (HPC) framework in conjunction with two public datasets of personal protective equipment (PPE) usage to categorize trainees’ poses into several clusters and visualize the average pose for each identified cluster.
The HPC framework was developed, which applies machine learning to video data to identify trainees’ reference action segments during PPE usage. To improve the clustering results, a skinned multi-person linear model was adopted to extract the three-dimensional (3D) coordinates of trainees’ body joints from a video frame and then create a 3D skeleton using the extracted joints. Also, various kinematic features and clustering algorithms were used to determine the best clustering method.
The significant findings were threefold. First, more similar poses were categorized into identical clusters as human poses were represented based on the distance between spine and other joints. Second, individuals’ representative action segments identified during PPE usage were different by the direction of motion as well as the location of key body joints. Third, the identified segments corresponded to significant behavioral steps for the correct PPE use.
This study presented a first look at how the type of kinematic features affects the performance of clustering PPE-related poses. In addition, this study provides theoretical contributions in that it clarified the difference in representative action segments in conjunction with their relevance to the correct PPE usage. Further, this study developed the HPC framework, which can identify reference segments of PPE-related actions.
