施工空间冲突事故险兆特征智能检测方法

Intelligent detection method for near-miss characteristics of construction space conflict accidents

为了更及时、准确地发现施工场景中潜在空间事故要素,减少相关事故发生。提出1种结合视觉检测与透视变换下的图像距离估计技术的施工空间冲突事故险兆特征检测方法,该方法以YOLOv7为基础模型,将主干网络替换为轻量级MobileOne,同时在特征融合网络与检测头中分别加入SPD-Conv卷积构建块及SimAM无参数注意力机制。实现目标检测、运动状态判别、距离测算3大功能模块的整合,感知空间冲突事故险兆特征进而及时阻断事故演进。研究结果表明:此方法在险兆目标检测任务中相较原始网络在mAP精度小幅降低2.79个百分点的情况下检测速度提升了50%;其次,运动属性险兆特征平均识别准确率达91%以上,测距精度保持在92%~96%。研究结果可为实现施工场景空间冲突险兆特征的有效检测提供参考。

In order to discover the potential spatial accident elements in the construction scene more timely and accurately,and reduce the occurrence of related accidents,a detection method for the near-miss characteristics of construction space conflict accidents combining with the image distance estimation technology based on visual detection and perspective transformation was proposed.Taking YOLOv7 as the basic model,the backbone network was replaced with lightweight MobileOne.At the same time,the SPD-Conv convolution building blocks and SimAM parameter-free attention mechanism were added to the feature fusion network and the detection head,respectively.The integration of three functional modules of target detection,motion state discrimination and distance measurement was realized,and the near-miss characteristics of space conflict accidents were perceived,so as to block the accident evolution in time.The results show that compared with the original network,the detection speed of this method is increased by 50%when the mAP accuracy is slightly reduced by 2.79%in the task of near-miss target detection.Secondly,the average recognition accuracy of motion attribute near-miss characteristics reaches above 91%,and the distance measurement accuracy is maintained in the range of 92%~96%.The research results provide a reference for the effective detection of spatial conflict near-miss characteristics in construction scenes.