基于全景可视化的综采工作面异常行为识别关键技术

Key technologies for identifying abnormal behaviors in fully mechanized mining faces based on panoramic visualization

针对煤矿工作面工人异常行为难以智能监管的问题,提出了一种基于OpenPose改进的轻量化人体姿态识别模型。采用MobileNet-v3替换原模型中的骨干网络,并采用较小的卷积核替代原7×7的卷积核,以提高模型检测速度并减小所占用资源;针对工作面摄像头拍摄范围有限,难以满足工作面全景可视化的需求,采用SIFT图像拼接算法,将多个监控视野相融合,实现工作面实时同步全场景监测的目的。试验测试结果表明:改进后的人体姿态识别模型准确率达到81.5%,占用内存为原模型的42%,而检测速度为原模型1.59倍;改进的人体姿态识别模型在减小占用内存的同时,大幅提高了模型检测速度并保持了较高的准确率,结合人体姿态识别与图像拼接算法能够实现对工人行为的实时监测。

A lightweight human pose recognition model based on OpenPose improvement is proposed to address the difficulty of intelligent supervision of abnormal behaviors of coal mine workers.The backbone network of original model is replaced with Mobile-Net-v3,and the original 7×7 convolutional kernels is replaced with a smaller convolutional kernel to improve model detection speed and reduce resource consumption.In addition,due to the limited shooting range of the working face camera,it is difficult to meet the needs of panoramic visualization of the working face.This article adopts the SIFT image stitching algorithm to fuse multiple monitoring perspectives and achieve real-time synchronization of full scene monitoring of the working face.After algorithm experiments,the improved human posture recognition model has an accuracy of 81.5%,occupies 42%of the memory of the original model,and the detection speed is 1.59 times that of the original model.The improved human posture recognition model not only reduces the memory consumption,but also greatly improves the speed of model detection and maintains a high accuracy.Combining human posture recognition and image stitching algorithm can realize real-time monitoring of workers’behaviors.