Research And Application On Cross-Border Recognition Method Of Key Area In Power Grid Based On YOLOv3

With the continuous development of my country’s power supply industry,some automated control technologies have gradually been applied,and the safety and efficiency of power supply work have been significantly improved.Among them,the power control automation room is the main place to realize automation technology,and its importance is self-evident.Therefore,it is very necessary to carry out reasonable monitoring of the computer room where these devices are stored.This paper conducts out-of-bounds detection for the computer room containing key equipment of the power grid.Compared with the traditional indiscriminate monitoring of the entire computer room,this paper can flexibly delineate the monitoring range according to the needs of the power grid department.The key monitoring range of different computer rooms is different,and the monitoring range is independently delineated by the power grid department near key cabinets.When the target is active in the designated area,the system remains silent,and when it leaves the designated active area and enters the unauthorized area,the system alarms.Generally,the delineated area defaults to a convex polygon.In this case,it is difficult to use traditional infrared monitoring instruments.After studying the application of traditional machine learning algorithms and deep learning algorithms in target detection,this paper focuses on the study of the representative machine learning algorithm HOG+SVM and deep learning algorithm YOLOv3 in the field of target detection.Finally,according to the specific problems faced by specific applications,a target detection network based on feature fusion is proposed on the basis of YOLOv3,and the following improvements are made:1)For the missed detection problem caused by small targets that are far away from monitoring and are affected by light and shadow,this paper specifically integrates HOG features that are not easily affected by light and shadow and can extract semantics into the small target detection unit of YOLOv3 to improve the accuracy of this dataset.detection accuracy.2)Aiming at the problem of missed detection caused by mutual occlusion between dense targets,this paper softens the original non-maximum suppression score function by means of linear weighting,so that the prediction frame with higher confidence score is properly retained,so as to reduce The purpose of small missed detection rate.3)In view of the problem that the anchor frame does not fit the target of the data set,this paper uses the K-means++ clustering algorithm to replace the original k-means clustering algorithm to obtain the anchor frame,and the data set in this paper is clustered before clustering.Targeted two-step screening is performed to eliminate extreme data,so that the anchor box generated after clustering is more in line with the size and aspect ratio of the data set in this paper.4)In response to the requirement of flexible delineation of monitoring range put forward by the power grid,this paper designs a reasonable algorithm for judging outof-bounds for the monitoring range of convex polygons,and realizes the judgment of whether the target crosses the boundary and intrudes into the autonomously delimited area through the algorithm.Finally,the effectiveness of the above improvements is verified by experiments,and on this basis,a border crossing detection system with surveillance video as input is implemented.The improved target detection algorithm not only ensures the real-time detection,but also improves the accuracy of the detection of out-of-bounds targets in key computer rooms.