Key Technology Research On Pilot Typical Abnormal Behavior Recognition Based On Deep Learning

Aviation safety is the eternal theme of the operation of the civil aviation system.According to accident statistics,60% to 80% of flight accidents are caused by human factors.Statistics from the Civil Aviation Safety Annual Report show that the proportion of flight accidents caused by pilots and crew factors is as high as 67.2%.With the rapid development of the civil aviation industry and the rapid increase in air transportation volume,higher requirements are placed on pilots.According to aviation accident statistics,the abnormal behavior of pilots is the main cause of flight accidents,and flight accidents and symptoms are directly or indirectly related to the abnormal behavior of pilots.How to effectively detect the pilot’s abnormal driving behavior,effectively regulate the possible consequences of the pilot’s abnormal behavior,and explore and establish an effective mechanism to reduce human error from the perspective of intrinsic safety has attracted the attention of many researchers.Therefore,it is of great practical significance to carry out the research on identification,detection and early warning of abnormal pilot driving behaviors to regulate pilots’ driving behaviors and ensure the safe operation of aviation.Based on the definition of the concept of abnormal pilot behavior,this paper selects the research objects of typical abnormal behavior of pilots playing mobile phones and smoking in the cockpit,collects video data of simulated cockpit pilot behavior,uses deep learning and computer vision methods to analyze the data,and uses target detection,Human skeleton recognition,ST-GCN network and other deep learning models,combined with software and hardware,developed a pilot abnormal behavior recognition prototype system,providing a reference for the establishment of risk intelligent monitoring and early warning driven by information data.The main research work includes:(1)Research on video data enhancement method of abnormal pilot behavior.A series of methods such as median filter noise reduction,Gamma correction,contrast enhancement,and histogram equalization are used to perform image enhancement and image noise reduction on the dataset of pilot abnormal behavior.(2)Propose an improved YOLOv4 pilot abnormal behavior recognition model.In view of the complex background of the aircraft cockpit,the problem of misidentification of pilots smoking and playing with mobile phones,an improved YOLOv4 algorithm model is proposed.After the backbone extraction network and upsampling,the attention mechanism module is added to solve the problem of gradient disappearance during training,and to improve the accuracy of small item inspection in complex backgrounds.Experiments show that the average accuracy of the improved YOLOv4 model is 6.66% higher than that of YOLOv4 model and 3.67% higher than that of YOLOv5.(3)A pilot attitude estimation algorithm based on the Open Pose model is proposed.Take advantage of ST-GCN network availability and superiority.This paper also uses model tuning experiments,when using a learning rate of 0.01,top-1 and top-5 have the highest accuracy.Compared with the existing methods,it also achieves better robustness and stronger generalization ability,which further improves the behavior recognition performance of the model in this paper.(4)Design and development of prototype system for identification of typical abnormal behavior of pilots.Using the front-end construction language of Py Qt5,the project was deployed to the Raspberry,and a pilot abnormal behavior recognition prototype system based on the fusion of improved YOLOv4 and Open Pose models was developed,which proved the pilot abnormal behavior recognition method and technical feasibility of this paper.