Research On Abnormal Human Behavior Recognition In Electric Power Operation

In power production and construction,the occurrence of power accidents will not only bring economic losses,but also cause personal injury.70-80% of the accidents in the power system are related to illegal production behavior of the staff.While improving personal quality of employees,behavior supervision is essential.Manual supervision has disadvantages such as resource waste and low efficiency.Abnormal behaviors can only be identified after viewing audio and video recordings,which has potential accidents and cannot meet current application requirements.Therefore,it is urgent to put forward a solution and adopt modern technical means to realize omni-directional,whole-process and real-time behavior supervision on production site.Power operation is usually carried out in the open air and complex outdoor environment.Therefore,characteristics of the power operation site are defined in this paper as the target is easily blocked,easily lost and easily affected by the environment.Based on this,this paper carried out research on abnormal human behavior recognition in electric power operation scenarios,mainly carrying out the following five aspects of work.In the aspect of image preprocessing,gray scale transformation is carried out on the image,and most of the interference impurities are removed by square equalization filtering,denoising and morphological processing,and good results are achieved.In terms of target detection,for the YOLOv3 network that is difficult to distinguish overlapping targets,prone to false detections,missed detections and slow detection speed,maintain the original advantages and improve it from four aspects: lightweight processing of the network,introduction of parameters GIOU in the improvement of the loss function,introduction of adaptive NMS algorithm in the frame selection optimization,and introduction of k-means++ in the clustering method can effectively remove external interference and obtain a relatively complete contour,improving the real-time performance and accuracy of the system.In terms of target tracking,aiming at the problem that it is difficult to search and locate again after the target is lost,the RT-MDNet tracking algorithm is introduced,which is combined with the improved YOLOv3 network,and the target loss discrimination mechanism is designed according to the cross-union ratio to achieve re-detection after target loss.and keep following.Experiments show that the tracking robustness of the algorithm is greatly improved,which can ensure that the operator is always within the detection range.In terms of skeleton feature extraction,in view of the complex structure of the Open Pose model and the easy loss of data points,the structure is simplified by sharing the convolutional layer.The missing point data is supplemented with the mean data of the before and after frames and incorrect bone connections are reduced by error correction,which improves the generalization ability of the model.In terms of behavior recognition,different from the previous algorithms that ignore the spatial relationship of human joints,the spatiotemporal graph convolutional neural network combined with clustering is used to identify abnormal behaviors.Through experimental verification and performance evaluation,the algorithm can meet the application needs of abnormal behavior recognition in power operation scenarios.