视觉联合交集下施工危险区域侵入行为动态感知方法

Dynamic Perception Method of Intrusion Behavior in Construction Dangerous Areas under Visual Joint Intersection

为解决行为险态识别脱离场景信息的问题,基于视觉联合交集,提出一种施工危险区域侵入行为动态感知方法。该方法以YOLOX模型为基础算法,引入轻量级Transformer,采用Mosaic数据增强方式,通过设计候选框联合交集(iou)将危险区域与工人行为状态关联,结合场景信息对工人行为险态进行理解。以临边作业为实验场景,在工人行为状态与危险区域位置关系精准识别的基础上,实现施工危险区域侵入行为的动态监控,改进后的YOLOX模型在每秒处理图像帧数降低1.33的情况下平均精度提高3.01%。实地测试结果表明:检测类别识别的平均精度均高于91.5%,iou准确率均为100%,可为施工场景内工人行为险态的实时感知提供参考,具有实效性与可推广性。

In order to address the issue of behavior risk identification separated from scene information,based on visual joint intersection,a dynamic perception method of intrusion behavior in construction dangerous areas is proposed.Based on the YOLOX model,the method introduces a lightweight Transformer,uses Mosaic data augmentation,and associates dangerous areas with behavioral states by designing a joint intersection of candidate boxes(iou) and understanding worker behavior risk status combined with scene informaiton.Taking the edge operation as the experimental platform,based on the accurate identification of location relationship between worker behavior status and dangerous areas,the dynamic monitoring of intrusion behavior in construction dangerous areas is realized.The improved YOLOX model achieves an average accuracy improvement of 3.01% with 1.33 fewer frames per second processed.The field test results show that the average accuracy of detection category recognition is higher than 91.5%,and the accuracy of iou is 100%,which provides a reference for real-time perception of worker behavior risk status in the construction scene,and has effectiveness and generalizability.