基于图卷积的作业行为实时检测方法

Real time detection method of work behavior based on graph convolution

针对复杂作业场景下的行为检测因光照、遮挡等因素导致错检、漏检且检测不达实时问题,提出一种基于图卷积的作业行为实时检测方法。对作业人员定位,过滤无关人员减少计算量,采用骨架提取并结合跟踪算法完成相邻帧间的目标匹配,采用优化后的图卷积神经网络进行不规范行为检测。以油田井场阀门操作为实验背景,该识别算法的处理速度能达到25fps,不规范阀门操作行为识别准确率达96.7%,能够达到实时识别。

Aiming at the problem of misdetection,missed detection and detection failure due to factors such as illumination and occlusion in behavior detection in complex operation scenarios,a real-time detection method of operation behavior based on graph convolution was proposed.The operator was located,the irrelevant personnel was filtered to reduce the amount of calculation,and the skeleton extraction and tracking algorithm were used to complete the target matching between adjacent frames.The optimized graph convolutional neural network was used to identify irregular behaviors.Taking the operation of valves in oilfield wellsites as the experimental background,the processing speed of the recognition algorithm can reach 25 fps,and the recognition accuracy of irregular valve operation behaviors can reach 96.7%,which can achieve real-time recognition.