Research On Fall Detection Method Based On Acceleration And Audio

In recent years,with the increasing trend of ageing in society,the health of the elderly has received a lot of attention from the community.As one of the main causes of death and injury of the elderly,falls have become the focus of medical health monitoring for the elderly.At present,single sensor and homogeneous sensor fall detection is not ideal,and heterogeneous sensor fusion fall detection is gradually becoming the mainstream of research.To this end,this paper uses acceleration signal data and audio signal data as the basis for fall detection research.The details of the research are as follows.First of all,based on the acceleration data,a normalized window interception algorithm(Anomaly location window algorithm,ALW)based on anomaly location is proposed to solve the problem of high resource consumption of the sliding window interception algorithm and low accuracy of subsequent fall detection.The algorithm uses a multicondition threshold judgment method to greatly reduce resource consumption on the premise of ensuring that no fall events are missed;the intercepted normalized window is further divided into impact stage and static stage,and features are extracted in stages according to the characteristics of each stage.Compared with the overall feature based on full-window extraction,the staged feature retains more information about the fall process,effectively improving the accuracy of fall detection.Secondly,based on audio data,aiming at the problems of poor robustness of fall detection based on audio data and sparse audio data of fall detection,a noisy data set mixed with background sound was constructed and a new audio detection algorithm with original audio waveform as input was designed.One-dimensional Residual Convolutional Network(1D-Re CNN).Compared with the fall detection model based on two-dimensional convolution,1D-Re CNN has a small number of parameters,which reduces the amount of data required for training,and avoids the process of two-dimensional spectral feature extraction and preprocessing of audio,and improves the model.Fall detection accuracy and efficiency in noisy environments.Finally,based on the acceleration data and audio data,a fall detection framework based on heterogeneous sensor fusion is designed.By analyzing the heterogeneity of acceleration data and audio data,the D-S evidence theory is used to fuse acceleration output and audio output at the decision level,which solves the problem of limited performance of single sensor and isomorphic sensor or high false alarm rate caused by the influence of environmental noise.The proposed fall detection model is evaluated in terms of sensitivity and false alarm rate by collecting simulated fall data from volunteers and free-living data in real-life environments.The experimental results show that the proposed model achieves a low false alarm rate of one false alarm every 26.67 hours on average in the 240-hour free life test while detecting all fall events.