Falls from height (FFH) are still a leading cause of fatalities in the construction industry, which also includes scaffolding-related accidents. Despite regular safety inspections, numerous scaffolding-related accidents occur at the construction site. The current safety monitoring practices are not only impractical but infeasible due to dynamicity of construction environment. Since a separate computer training and detection process is generally required to acquire spatiotemporal reasoning to control a single hazard; thus previous efforts in vision intelligence applications to improve safety monitoring are still limited to specific hazards. Also, in regard to detecting unsafe situations based on extracted correlations from safety rules, to date, previous studies have devoted little attention to this domain. To address these issues, this study proposes a correlation-based approach for mobile scaffold safety monitoring and detecting worker's unsafe behaviors. A deep neural network, Mask R-CNN, was used as classification and segmentation of worker's tasks combined with object correlation detection (OCD) module to identify worker's unsafe behaviors. The approach divides the overall construction worker's safety into two subsets, classification of worker and detection of safe (class-1) and unsafe (class-2) behavior using OCD block. The overall performance was evaluated on set of real scenarios with test results showing 85 % and 97 % precision and recall for class-1 (safe behavior) and 91 % and 65 % precision and recall for class-2 (unsafe behavior). The overall accuracy of 86 % confirms the Mask R-CNN-based OCD module's applicability for detecting worker's unsafe behavior effectively in a construction environment.

高空坠落（FFH）仍然是建筑行业死亡的主要原因，其中还包括与脚手架相关的事故。尽管定期进行安全检查，但在施工现场仍发生许多与脚手架有关的事故。由于施工环境的动态性，当前的安全监控实践不仅不切实际，而且不可行。由于通常需要单独的计算机培训和检测过程来获取时空推理，以控制单个危害；因此，因此，以前在视觉智能应用中为改善安全监控所做的努力仍然仅限于特定的危害。另外，到目前为止，关于基于从安全规则中提取的相关性来检测不安全情况的方法，以前的研究很少关注此领域。为了解决这些问题，这项研究提出了一种基于相关性的方法来移动脚手架安全监控和检测工人的不安全行为。深度神经网络Mask R-CNN被用作工人任务的分类和分段，并与对象相关检测（OCD）模块相结合，以识别工人的不安全行为。该方法将整体建筑工人的安全分为两个子集：工人的分类以及使用OCD块检测安全（1级）和不安全（2级）行为。在一组真实场景中评估了总体性能，测试结果显示1级（安全行为）的准确度和召回率分别为85％和97％，2类（不安全行为）的准确度和召回率分别为91％和65％。86％的总体准确度证实了基于Mask R-CNN的OCD模块在检测工人是否受伤方面的适用性。