A Voxel-Based 3D reconstruction and action recognition method for construction workers

Abstract

Workers are critical yet unpredictable elements on construction sites, with their actions significantly impacting safety and productivity. Recognizing these actions is essential for improving efficiency, safety, and quality. Benefiting from the advantages of the voxel format in terms of universal representation, privacy protection and memory saving, a voxel-based method for action recognition was proposed. By transforming the image into a structured voxel, a lightweight 3D CNN, CVARnet (Construction worker Voxel Action Recognition network) was established. To verify the effectiveness of voxel and CVARnet, a dataset named Construction Action Voxel Classification (CAVC) was developed. The image was primarily sourced from the construction site and represented five types of typical actions. The proposed CVARnet achieved 86% ACC in a classification task, demonstrating efficient recognition capabilities for workers’ actions. This study presented a novel perspective with a voxel format, providing innovative insight for the action recognition task.