Research On Monitoring Target Fall Detection Algorithm Based On Deep Learning

In recent years,as the aging of China's population continues to accelerate,the health,protection and rescue of the elderly are receiving a great deal of attention from society.Falls are one of the most common problems among the elderly and the disabled,especially those living alone who are unable to get help in time after a fall and pose a threat to life and health.Therefore,it is necessary to research a fall detection algorithm.In the field of intelligent monitoring,fall detection algorithms based on deep learning have become a hot research direction.Specifically,it uses a convolutional neural network to detect the main points of the human body and extract posture information.This is an important prerequisite for getting the characteristics of a falling action,however,the multi-scale issues of the target,the variety of human poses,and the interference of complex scene factors cause pose occlusion and loss,which is not useful for extracting the characteristics of the fall,and it is also susceptible to similar behavior when detecting falls,which makes algorithm research more difficult.Considering the above issues,the thesis researches on the monitoring target fall detection algorithm based on deep learning,and the design idea of the fall detection algorithm is adopted in the two-step method of posture estimation and feature extraction and classification to improve the accuracy of fall detection algorithm.The main research contents are as follows:1)Aiming at the problem that the detection accuracy of key points was low due to the multi-scale target,the diversity of human posture and the complex scene in the surveillance video,a human pose estimation model based on the improved R-FCN was constructed.A top-down body posture estimation structure was used to obtain the target posture information.The R-FCN model was selected as the basic architecture to build the target detection module,and the original Res Net basic network was changed into the Res Ne Xt network with better performance and higher efficiency,and the multi-scale candidate region module was improved,and human skeleton diagram was obtained by combining the key point detection network and One-Hot coding technology.In the test of the public data set,the average detection accuracy of key points of the human body reaches 94.6%.The experimental results show that the model effectively solves the multi-scale problem of the target,and obtains better results for pose estimation in situations where poses are difficult and the scene is complex.At the same time,compared with other models,it improves the detection speed.2)To solve the problem that fall detection was easily disturbed by similar behaviors and background factors,a spatiotemporal graph convolutional network model based on attention mechanism was proposed.The model used the advantages of graph convolutional network to extract features from spatial dimension and the advantages of gated recurrent unit to extract features from temporal dimension to obtain spatial and temporal features of skeleton sequences respectively,and introduced the attention mechanism to optimize the performance of the network and to classify and predict falling actions and non-falling actions.In addition,before extracting features,a key point complement algorithm was used to complement the missing skeleton to improve the integrity of the skeleton diagram.In the test of the public data set,the sensitivity of fall motion detection reaches 99.5%.The experimental results show that the model has obvious excellent performance under the interference conditions of complex background and target occlusion,scale change and similar behavior.In summary,through the study and discussion of the relevant theories of research contents,the plan is optimized for actual needs,the monitoring target fall detection algorithm based on deep learning studied in the thesis is proved by experiments: the detection speed of the monitoring system can basically meet the needs of the actual application and reduce the false alarm rate to some extent,compared to other fall detection algorithms,the accuracy of fall detection is improved.