Abstract

The construction industry is among the world's most dangerous industries, with a high number of accidents and fatalities. Following safety guidelines and wearing the required Personal Protective Equipment (PPE) is an essential step in mitigating accidents. Safety managers and inspectors are responsible for making sure safety regulations are correctly followed. However, safety inspection is time-consuming, costly, and is done based on a random basis and for a short period.
In order to facilitate safety inspection, various research studies are done using different techniques and technologies. Detecting PPE using Computer Vision (CV) has gained a lot of interest in enhancing construction sites' safety. Nevertheless, detecting PPE on large construction sites and generating safety reports is still a big challenge. Additionally, real-world 2D localization of workers is critical to monitor workers’ safety based on their location. This research proposes an automated framework consists of three modules to enhance the safety of construction sites.
The first module of the framework is the PPE Detection (PPED) module, which detects and tracks the workers and their PPE on large construction sites based on the frame segmentation technique. The second module is the PPE Safety Report Generation (PPESRG), which uses PPED results to match workers in two overlapping views and generate technical and practical high-level safety reports while protecting workers’ privacy. Finally, the third module of the framework is a Single-camera Localization (SL) module that uses worker detection results from the PPED module and camera calibration parameters to locate workers on 2D real-world coordinate and monitor workers’ safety based on their location on the construction site.
 
The proposed framework is validated using real-world construction videos, and the experimental results of each module demonstrate the practicality and robustness of applying on real-world construction sites. Based on different test videos, the PPED module has achieved 99.04% precision, 91.61% recall, and 90.77% accuracy. Furthermore, the generated safety reports are validated by the safety managers of the project as being practical for safety monitoring on the construction sites. Finally, the proposed CL module is validated with an average error of the average 1.58 m for locating workers on the construction sites.
The main contributions of this research are: (1) proposing a nested network based on frame segmentation technique that improved the worker and PPE detection rate on large construction sites, (2) proposing a safety report generation method, which benefits from PPED results of two cameras to generate practical safety reports while protecting workers’ privacy and (3) single-camera based technique which is fast and easy to implement on large construction sites in order to locate workers. Future works will focus on accelerating the detection process, improving CV-based localization accuracy, and benefitting from other data sources to enhance generated safety reports (e.g., schedule, etc.).