Deep learning for site safety: Real-time detection of personal protective equipment

摘要

The leading causes of construction fatalities include traumatic brain injuries (resulted from fall and electrocution) and collisions (resulted from struck by objects). As a preventive step, the U.S. Occupational Safety and Health Administration (OSHA) requires that contractors enforce and monitor appropriate usage of personal protective equipment (PPE) of workers (e.g., hard hat and vest) at all times. This paper presents three deep learning (DL) models built on You-Only-Look-Once (YOLO) architecture to verify PPE compliance of workers; i.e., if a worker is wearing hard hat, vest, or both, from image/video in real-time. In the first approach, the algorithm detects workers, hats, and vests and then, a machine learning model (e.g., neural network and decision tree) verifies if each detected worker is properly wearing hat or vest. In the second approach, the algorithm simultaneously detects individual workers and verifies PPE compliance with a single convolutional neural network (CNN) framework. In the third approach, the algorithm first detects only the workers in the input image which are then cropped and classified by CNN-based classifiers (i.e., VGG-16, ResNet-50, and Xception) according to the presence of PPE attire. All models are trained on an in-house image dataset that is created using crowd-sourcing and web-mining. The dataset, named Pictor-v3, contains similar to 1,500 annotated images and similar to 4,700 instances of workers wearing various combinations of PPE components. It is found that the second approach achieves the best performance, i.e., 72.3% mean average precision (mAP), in real-world settings, and can process 11 frames per second (FPS) on a laptop computer which makes it suitable for real-time detection, as well as a good candidate for running on light-weight mobile devices. The closest alternative in terms of performance (67.93% mAP) is the third approach where VGG-16, ResNet-50, and Xception classifiers are assembled in a Bayesian framework. However, the first approach is the fastest among all and can process 13 FPS with 63.1% mAP. The crowed-sourced Pictor-v3 dataset and all trained models are publicly available to support the design and testing of other innovative applications for monitoring safety compliance, and advancing future research in automation in construction.