基于预测-重构融合的飞行员异常驾驶行为检测方法研究OA

Abnormal pilot driving behavior is one of the main causes of aviation accidents.Therefore,accurately detecting the starting and ending moments of abnormal pilot behavior is of great significance to improving flight safety.In this paper,a prediction-reconstruction fusion method is proposed for abnormal pilot driving behavior detection.This method consists of two modules,namely future frame prediction module and current frame reconstruction module.The former aims to predict future frame based on past frames,which is implemented via a spatial-temporal UNet.The latter focuses on reconstructing current frame,which is achieved through a diffusion model.Specially,current frame reconstruction module generates current frame conditioned on latent features in future frame prediction module,which can constrain the feature extraction process of the future frame prediction module,thereby enhancing the ability to understand both global and local features of driving behavior.Furthermore,this paper proposes a personalized score fusion inference strategy,which can comprehensively detect abnormal behaviors by integrating motion information from the future frame prediction module and appearance information from the current frame reconstruction module.Extensive experiments on the publicly available CUHK Avenue dataset and the modified Driving-VAD dataset demonstrate the effectiveness of this method.The proposed method achieves detection accuracies of 91.3％and 89.9％on these two datasets,respectively,outperforming other existing methods.Additionally,the model parameter sizes for the future frame prediction-only model and the prediction-reconstruction model are 0.4 GB and 1.62 GB,respectively,with average inference times of 0.4 s and 2.9 s per video frame,demonstrating a certain potential for practical applications.This research not only helps standardize pilot operations but also plays a crucial role in ensuring aviation safety.