Research On Dangerous Action Recognition Method Of Personnel In Petrochemical Scene Fused With Object Detection

The petrochemical industry(hereinafter referred to as "petrochemical")is a highrisk industry.In this scenario,it is strictly prohibited to use open flames,cellphone calls,and smoking that can easily cause explosions.It is necessary to identify and warn such behaviors in a timely and accurate manner.At present,there are not many researches on the identification of dangerous behaviors such as phone calls and smoking in petrochemical scenes.Traditional methods mainly analyze surveillance videos through on-site persuasion or manual analysis.However,manual supervision is very prone to untimely or oversight,and it is difficult to identify and deal with dangerous behaviors in time.At the same time,traditional human action recognition algorithms only focus on personal behaviors,and cannot identify mobile phone calls,smoking,etc.,which are human-object interactions.The dangerous behaviors cannot meet the ever-increasing security monitoring needs in the petrochemical scene.Aiming at the above practical problems,a method for identifying dangerous behaviors of personnel in petrochemical scenes with fusion of object detection is proposed.The main work of this thesis is as follows:On the one hand,in view of the low detection accuracy and low recall rate of small objects such as cell phones and cigarettes in surveillance videos by existing object detection algorithms,this thesis improves the multi-scale and clustering algorithm based on the YOLOv3 algorithm,and proposes a multi-scale-YOLOv3(MultiscaleYOLOv3,M-YOLOv3)object detection algorithm.First,in order to make full use of the rich feature information of small objects in the shallow network,the original 3-scale prediction is extended to 4-scale to improve the detection ability of small objects;secondly,in order to overcome the original K-means algorithm in the YOLOv3 algorithm The clustering results are related to the initial point selection,strong randomness and other shortcomings.The K-means++ algorithm is used to optimize the anchor box selection strategy,improve the average intersection ratio between the real anchor box and the predicted anchor box,and improve the recall rate;finally,a object detection data set containing two categories of cell phones and cigarettes is established and experimentally verified on this basis.Experimental results show that the precision and recall rate of the M-YOLOv3 object detection algorithm proposed in this thesis have been greatly improved.On the other hand,in order to improve the recognition effect of human-object interaction behaviors such as using cell phones and smoking,this thesis introduces the object detection mechanism into the skeleton based human action recognition task,and proposes a method of human-object interaction action recognition in petrochemical scenes fused with object detection.First,the Open Pose algorithm is used to obtain the skeleton information of people in the video,and then the action recognition method is used to obtain the initial action category;second,the M-YOLOv3 algorithm proposed in this thesis is used to detect the object of interest to obtain its category and location information;Then characterize the human-object interaction relationship by judging the spatial relationship between human and object;Finally,propose a decision-making fusion strategy,which merges the initial action category of the human,object information,and human-object interaction relationship to obtain the final action recognition result.Collecting and establishing a cell phone calls and smoking action data set.The experimental results show that the performance of the improved human action recognition method has been greatly improved,and it can accurately recognize the dangerous of people in the petrochemical scene.The thesis has 47 pictures,14 tables,and 77 references.