Detection of construction workers under varying poses and changing background in image sequences via very deep residual networks

초록

Analyzing the location and behavior of construction workers using construction site images has been recognized as a means of providing useful information for safety management and productivity analysis. Although effective utilization of analyzed image data requires accurate and timely detection of workers in complex, continuously changing working environments, the previous methods that detect construction workers still require improvement because of the poor detection performance. This study proposes the use of very deep residual networks to accurately and rapidly detect construction workers under varying poses and against changing backgrounds in image sequences. The architecture of construction worker detection in this study is based on convolutional neural networks (CNNs). The proposed method is divided into two stages: extracting feature maps via very deep residual networks (ResNet-152) and bounding box regression and labeling from the original image via Faster regions with CNN features (R-CNN). The experiments were conducted at actual construction sites by acquiring 1.3-megapixel and 3.1-megapixel images from a movable digital camera to verify the proposed method for images from fixed and moving cameras. Faster R-CNN with ResNet-152 had accuracy, precision, and recall rates of 94.3%, 96.03%, and 98.13% for 3241 images, respectively. The proposed method processed 0.2 s per frame (i.e., 5 frames per second) on average. The results show that it is possible to accurately and rapidly detect multiple workers in construction site images by employing very deep residual networks without relying on limited assumptions about workers' postures, appearance, and background.

키워드