Inspection Behavior Detection Of Underground Distribution Room Based On Weakly Supervised Learning

In recent years,the safety situation of coal mine production in my country has been improving day by day,but major accidents still occur frequently,and long-term stable power supply is a necessary condition for coal mine production safety.In this regard,it is necessary to ensure high-quality daily inspection of the underground power distribution room,so as to detect and deal with the existing safety hazards in time.The research focus of the existing underground power distribution room inspection behavior detection method is the video action classification based on fully supervised learning,which has the problems of cumbersome data set production and inability to locate the action sequence.Therefore,this paper proposes to use the action detection method based on weakly supervised learning to detect the inspection behavior in the underground power distribution room,so as to ensure that the inspectors conduct inspections according to the inspection rules and ensure the safety of the coal mine power supply system.However,in the weakly supervised learning action detection task,there are problems that the action frame and the background frame are easily confused and the predicted action instance is incomplete,which makes the detection accuracy low.Aiming at the above problems,this paper designs an action detection model based on weakly supervised learning,which effectively improves the detection performance of the model.At the same time,the model is applied to the inspection and detection task of the underground power distribution room,which can classify and locate the inspection actions in the surveillance video,which effectively improves the supervision efficiency of the inspection work.The specific work is as follows:(1)This paper proposes a point-supervised action detection model based on conditional variational autoencoders.First,the attention module is used to extract the attention of actions and backgrounds in the features.Then,a generative model is used to facilitate learning the potential differences between action frames and background frames,which effectively separates action frames from background frames.Finally,the concept of external-internal contrast is introduced,and an optimal sequence evaluation method is proposed,which improves the integrity of predicted action instances.On the THUMOS14 dataset,m AP@0.5 reaches 41.3%,which is 10.8% higher than the current similar methods.(2)A data set of downhole power distribution room inspection behavior is created,and verify the model performance on this data set.First,using the monitoring video of the power distribution room,the detection behaviors are divided into inspection standing,inspection squatting,walking,recording standing and recording sitting,and a data set of inspection behaviors of underground power distribution rooms is created and constructed.Second,the feature extraction model is used to extract the RGB features and optical flow features of the monitoring video of the underground power distribution room,and then the obtained RGB features and optical flow features are input into the pointsupervised action detection model based on conditional variational autoencoders for training.After the training is completed,the model can output the time interval,action label and confidence level of the patrol action in the surveillance video.The experimental results show that the model can simultaneously complete the task of inspection behavior classification and localization,and the m AP@0.5 on the self-made inspection behavior data set reaches 46.8%.