| The power scene has the characteristics of wide spatial range,many personnel,complex operation behavior,and high-altitude operation,and safe operation plays a decisive role in this scene.In recent years,intelligent video analysis technology represented by object detection and attitude estimation has been gradually used in the safety monitoring of the power industry to ensure the safe operation of live operations.Based on this,this paper focuses on the intelligent safety monitoring method of live operators,and studies from three aspects: safety protective equipment detection,operator attitude estimation,and safe ranging measurement between human body and charged body in three-dimensional space,and the work is as follows:1)Safety protection equipment detection for live workers.In power scenarios,protective equipment such as insulating poles and blankets are large in length and sensitive to angular changes during object detection.In this paper,the CSL-based object detection algorithm is used to complete the object detection task with a rotating frame,and then determine whether the live operator wears the safety protective equipment correctly.The average detection accuracy of all targets in the collected power data set is more than95%.2)Estimation of the attitude of live operators.The bottom-up hyperpose pose estimation library was used to monitor the safety behavior of live workers in working scenarios.The attitude estimation algorithm is used to obtain the relative position relationship between key points to determine whether the operator’s behavior is dangerous.The pose estimation model can reach 47.3m AP.3)Safe distance measurement between live operators and charged bodies.By fusing the target detection results and the three-dimensional point cloud obtained by lidar,the position information of two-dimensional target points in three-dimensional space is obtained,and then ranging is completed.The original point cloud data is very sparse,and when the two-dimensional target point lacks the corresponding three-dimensional information,the improved MSG-CHN depth map is used to complete the network to predict the three-dimensional coordinates of the target point.Aiming at the problem of sparse and irregular distribution of point cloud data,sparse invariant convolution is added to the deep branch of the original network,and the mask is used to predict the sparse depth feature and confidence degree at each layer of convolution,so as to reduce the input sensitivity of the network to different sparsity degrees.Aiming at the problem that simple manual down-sampling used in RGB branch does not sufficiently guide depth prediction,the attention mechanism is added in this branch to improve the feature extraction function of image edge and semantics.Experiments on the kinect＿leap and KITTI datasets show that the combination of attention mechanism and sparse invariant convolution can obtain optimal prediction results at 3.46 M model size.The improved algorithm has an accuracy of 130 mm and 801 mm in the two data sets,which is 26 mm and 117 mm higher than the original network,respectively.4)Reconstruction of the hands of live workers.In this paper,the completion effect is observed by real-time reconstruction of the manpower information after depth map completion.The hand contour information after object detection is obtained in real time through the Mobrepon algorithm,and the depth map corresponding to the contour part is projected back to the 3D point cloud to complete the real-time reconstruction based on the PCL library.In intelligent safety monitoring,it is necessary to return the safety status of the job site in real time,which puts forward higher requirements for the computing power of algorithms and edge devices.This paper completes the above algorithm based on the UAV platform,and analyzes the safety status of live workers in real time.Once violations are found,audio equipment can be used to issue an alarm immediately.After testing,the monitoring system can infer that the speed can reach 25 FPS when the target detection and attitude estimation are turned on at the same time,and the safe distance measurement can reach 15 FPS when combined with the laser radar.The speed of 3D reconstruction is10 FPS. |
