基于边云协同框架的煤矿井下实时视频处理系统

Real-time video processing system in coal mine based on edge-cloud collaborative framework

目前煤矿井下智能视频监控主要采用云计算方式处理实时视频,视频传输占用的网络资源多,时延高,无法实时响应监控区域发生的紧急事件。针对该问题,提出了基于边云协同框架的煤矿井下实时视频处理系统,将实时性强的目标识别任务下放至边缘端,将计算量大且实时性弱的边缘设备整合等任务放至云端处理。在视频监控现场,利用部署在边缘设备上的神经网络模型对视频监控图像进行本地处理;通过井下异构融合网络将不同网络环境中边缘设备的处理结果和模型参数等信息发送给云服务器;云服务器针对性地对各场景中的边缘设备进行模型更新、推送,最终实现边云数据实时交互和边缘端服务的在线优化。针对目标检测模型Tiny-YOLOv3无法提取到图片的深层特征、易出现梯度消失和过拟合现象等问题,依据残差结构设计了下采样残差模块,对Tiny-YOLOv3进行改进,以提高模型的深度特征提取和泛化能力。在边云数据交互的基础上,对边缘设备上的目标检测模型进行针对性场景优化,以提高边缘设备端模型检测的准确率。测试结果表明:改进型Tiny-YOLOv3模型的稳定性与数据泛化能力优于YOLO和Tiny-YOLOv3;经过单一场景的特化训练后,改进型Tiny-YOLOv3模型的目标识别更加精准;与云计算相比,边云协同框架可显著降低监控视频处理时延。

At present,the intelligent video monitoring in coal mine mainly adopts cloud computing to process real-time video,and the video transmission occupies large network resources and has high delay,which can not respond to emergency events in the monitoring area in real time.In order to solve this problem,a real-time video processing system in coal mine based on edge-cloud collaborative framework is proposed.In this system,the real-time target recognition task is sent to the edge,and the tasks with large calculation and weak real-time performance such as edge device integration are sent to the cloud for processing.At the video monitoring site,the neural network model deployed on the edge device is used to process the video monitoring image locally.Through the underground heterogeneous fusion network,the processing results and model parameters of the edge devices in different network environments are sent to the cloud server.The cloud server updates and pushes the model of edge devices in each scene,and finally realizes real-time interaction of edge-cloud data and online optimization of edge services.In order to solve the problems that Tiny-YOLOv3 can not extract the deep characteristic of the image,and is prone to gradient disappearance and over-fitting,a down-sampling residual module is designed according to the residual structure,and Tiny-YOLOv3 is improved to improve the deep characteristic extraction and generalization capability of the model.On the basis of edge-cloud data interaction,the target detection model on the edge device is optimized for the targeted scene to improve the accuracy of model detection on the edge device.The test results show that the stability and data generalization capability of the improved Tiny-YOLOv3 model are better than those of YOLO and Tiny-YOLOv3.After specialized training in a single scene,the improved Tiny-YOLOv3 model is more accurate in target recognition.Compared with cloud computing,the edge-cloud collaborative framework can reduce the latency of monitoring video processing significantly.