构件吊装作业是装配式建筑施工中最为频繁，也是较为危险的一项作业，特别是物体打击等安全事故频发。目前，针对装配式建筑施工的安全管理研究与实践，主要集中在风险量化分析和现场安全评估上，缺乏对施工过程安全风险的事中监控，也没有聚焦到具体的作业场景中。同时，愈发成熟的物联网以及计算机视觉技术为施工现场安全监控提供了一种可行的解决思路，可以有效替代传统的人工检查方式。然而，现有相关研究与实践大多集中于对施工现场单个不安全场景的监控，没有很好地考虑作业过程的连续性和场景之间的逻辑关系，特别是较少考虑装配式建筑施工场景的安全监控。 为此，本研究将聚焦于装配式建筑构件吊装安全，提出其智能监控方法。首先，基于装配式建筑构件吊装流程及特点，系统地分析了吊装作业安全风险以及相关标准中对应的安全管理要求，并从事中监控的角度出发选取不安全场景作为监控对象。然后，对吊装作业进行了场景划分和定义，并提出连续检测框架构建要求，以预制墙板和预制叠合板为例，分别构建了其不安全场景连续检测框架。进而，针对连续检测框架中的具体不安全场景，建立了基于计算机视觉的智能识别方法。最后，以预制墙板构件支护以及落吊场景为例，对上述不安全场景识别方法进行了验证分析。结果显示：1）在支护安全性识别测试中，所构建的预制墙板及斜支撑对象识别模型的平均精度均值达到95.3%，基于该模型建立的支护安全性识别方法的准确率为70.7%，检测速度约为10fps。2）在落吊场景安全性识别测试中，墙板相对于作业面高度计算以及工人与构件水平距离计算误差分别为35.4mm和112.2mm，墙板对象识别丢失率为16.9%，落吊场景安全性识别的准确率和在墙板被成功识别条件下的准确率分别为77.7%和93.5%，检测速度约为10fps。上述测试结果表明，本研究所提出的不安全场景识别方法是有效的，在距离计算精度、检测实时性以及场景识别准确性上均能够满足施工现场的需求。 本研究所提出的安全监控方法为吊装过程中多场景的持续安全监控提供了思路，也可以为其他作业流程的安全监控提供参考，丰富了施工安全管理理论与方法。同时，该方法能够有效识别构件吊装过程中的不安全场景，为现场的安全管理决策提供依据，减少施工安全事故发生。因此，本研究具有较好的理论意义与实践价值。
The hoisting of components is the most frequent, and very dangerous operation during prefabricated construction, in which safety accidents such as stuck by object happen frequently. At present, the research and practice in the safety management of prefabricated construction mainly focuses on the quantitative analysis of risks and on-site safety level assessment, but lacking the monitoring of safety risks in construction processes, and less focusing on specific operation scenarios. Meanwhile, the rapid development of Internet of Things and computer vision technology provides a powerful solution for on-site safety monitoring, this can effectively replace the traditional manual inspection. However, existing research and practice mainly focuses on the monitoring of single unsafe scene in construction sites, less considering the continuity of operation processes and the logical relationship between different scenes, particularly the safety monitoring of prefabricated construction scenes. This research aims at proposing an intelligent approach to monitoring safety in hoisting prefabricated building components. Firstly, based on the processes and features of prefabricated building components hoisting, this research systematically summarizes the potential safety risks and corresponding safety management requirements in standards, and selects the unsafe scenes as monitoring objects from the perspective of in-process control. Then, the hoisting processes are divided into scenes and well defined, and the continuous detection frameworks of unsafe scenes of precast wall and precast superposed slab are established respectively as examples with the establishing requirements proposed. Furthermore, aimed at the recognition of specific unsafe scenes in the continuous detection framework, the computer vision-based intelligent recognition methods of unsafe scenes are established. Finally, vertical support erection and precast wall hoisting are taken as examples to verify the recognition methods above. The results show that: 1) In the test of supporting scene safety recognition, the mean Average Precision of the model for detecting precast walls and inclined supports is up to 95.3%, the accuracy of the recognition method is 70.7%, and the detection speed is approximately 10fps. 2) In the test of hoisting scene safety recognition, the errors of the wall elevation calculation over the operation surface as well as the horizontal distance calculation between workers and components are 35.4mm and 112.2mm respectively, and the miss rate of the wall object is 16.9%; The overall accuracy of the recognition method and the accuracy under the condition of wall being successfully detected are 77.7% and 93.5% respectively, and the detection speed is around 10fps. In summary, the above test results indicate that the unsafe scene recognition methods are effective, being able to meet practical requirements in construction sites in terms of distance calculation accuracy, real-time detection and unsafe scene recognition accuracy. This intelligent approach to monitoring safety in hoisting prefabricated components provides not only a new idea for realizing multi-scene persistent safety monitoring during the process of hoisting, but also a valuable reference for the safety monitoring of other operation processes, thus enriching the theory and method of construction safety management. Moreover, this method can effectively recognize the unsafe scenes during the process of component hoisting, thus providing supports for safety management decision and reducing construction safety accidents. Hence, this research is with the significance of both theory and practice.
