Treffer: Convolutional neural networks for construction safety: A technical review of computer vision applications.

The increasing number of construction-related accidents underscores the urgent need to enhance global construction safety (CS) management. Monitoring activities during the construction process is essential for effective CS management. Traditional computer vision (CV) techniques, which rely on handcrafted features and rule-based algorithms, often struggle to capture the complex dynamics of construction sites. In contrast, deep learning-based CV approaches, especially convolutional neural networks (CNNs), offer end-to-end solutions that overcome these limitations and are increasingly adopted in CS applications. This paper systematically categorizes the application of CNN-based CV technologies into four key stages: data collection, data preprocessing, model construction, and practical application. From a technical perspective, this paper provides a comprehensive review of relevant methods and tools at each stage. Initially, this paper analyzes the literature on CNN-based CV applications in CS via the Python library pyBibX combined with artificial intelligence (AI) tools for bibliometric analysis and effective visualization. This paper subsequently summarizes various data acquisition and preprocessing methods to save researchers time in data-related aspects. Furthermore, to provide researchers with a rapid understanding of existing methods, this paper presents various CV and CNN techniques, summarizing classical CV models based on CNNs through extensive literature, data, web, and competition searches, and compares their performance on public datasets. Finally, this paper provides a detailed analysis of CNN-based CV technologies used in CS, tailored to the technical aspects of various downstream tasks, revealing current applications of advanced technologies and research progress. Additionally, this paper discusses current research challenges, including technical and other aspects, and proposes several directions for future research. [Display omitted] • Four-stage framework: data collection, preprocessing, model construction, and practical application. • Bibliometric analysis of CNN-based CV in CS using PyBibX and AI tools. • Summary of data acquisition and preprocessing methods for CNN applications. • Comparison of performance of classical CNN models on public datasets. • In-depth analysis of CNN-based CV algorithms for various CS tasks. [ABSTRACT FROM AUTHOR]