基于机器视觉的施工人员危险行为监测与识别技术——以某矿坑公园为例

Construction Worker Hazardous Behavior Monitoring and Recognition Technology Based on Machine Vision—A Case Study of a Certain Quarry Park

摘要: 超过80%的建筑安全事故源于施工人员的危险行为，而传统人工管理方式的滞后性限制了监测效果，迫切需要引入智能算法以提高监测精度和实时性。本文首先通过事故报告的统计数据对施工危险行为的进行了评估和分类。以某矿坑公园为实证案例，提出了一套监测点优化布置方案。在此基础上，采用YOLO v5和CNN-LSTM算法实现了危险行为的精准识别。研究结果表明，监测点布置方案经优化后，各覆盖参数均取得不同程度的提升，监测覆盖率提高了2.17%，崖壁监测覆盖长度增加了40.33%。在状态类行为识别方面，采用YOLO v5算法实现了高达97.2%的识别精度。对于动作类行为，引入CNN-LSTM算法，其准确率达到95.8%，视频帧数达到30帧/s，满足实时监测需求。该研究在实际应用场景中为施工人员危险行为监测提供了有效的技术支持。

Abstract: More than 80% of the construction safety accidents are caused by the hazardous behavior of construction personnel, and the lag of traditional manual management limits the monitoring effect. To improve the monitoring accuracy and real-time performance, there is an urgent need to introduce intelligent algorithms. First of all, this paper evaluates and classifies the construction hazardous behavior through the statistical data of the accident report. Taking a certain quarry park as an empirical case, a set of optimal layout scheme of monitoring points is put forward. On this basis, the accurate identification of hazardous behavior is realized by using YOLO v5 and CNN-LSTM algorithm. The results show that after the layout of the monitoring site is optimized, the coverage parameters are improved in varying degrees, the monitoring coverage rate is increased by 2.17%, and the cliff monitoring coverage length is increased by 40.33%. In the aspect of state behavior recognition, the recognition accuracy of 97.2% is achieved by using YOLO v5 algorithm. For action behavior, the CNN-LSTM algorithm is introduced, and its accuracy is 95.8%, and the number of video frames can reach 30 frames/s, which meets the needs of real-time monitoring. This study provides effective technical support for hazardous behavior monitoring of construction workers in practical application scenarios.