Research On Fall Detection And Pre-impact Warning Technology Based On Multi-sensor

In the context of an aging global population,resulting in more fall accidents in the elderly.In recent years,it has attracted considerable research attentions.In order to reduce the injuries caused by falls,related researchers have tried to detect and predict the fall with scientific means.The advancement of micro-sensor and integrated technology provide a promising tool for obtaining valuable data related to human motion status.The unique advantages of micro-sensor,together with the application of specific algorithms,make the user's motion status can be detected.However,most of existing works adopt the single-sensor based detection architecture and fail to distinguish falls from fall-like activities.Therefore,this paper is committed to explore multi-sensor signal processing schemes,and design a multi-sensor-based framework to solve human fall detection and pre-impact warning problems,so as to further improve detection accuracy and prediction real-time performance.The main works and contributions are summarized as follows:(1)By analyzing the characteristics and functions of smart insole self-designed,the data collected by three sensors embedded in it is used as the feature information of human fall.Apply the one-dimensional convolutional neural network(1D CNN)to the fall detection system,and use the collected acceleration data and velocity data for training and testing.Compare with traditional machine learning methods(such as SVM,KNN),the proposed method shows its superiority in accuracy,sensitivity and specificity.(2)We utilize CNN to detect the fall direction in the pre-impact manner.To further distinguish the direction of falls that most existing studies have not considered,we propose a multi-sensor-based fall detection system by taking the detection as a multi-class problem.The direction of falls can be used for activating an appropriate protection device and reducing the severity of traumatic injury.(3)A Multi-source CNN ensemble structure called MCNNE is proposed to improve the overall detection accuracy.In the proposed system,data from different sensors are preprocessed and formatted as the training dataset independently,and output features map from different sensors are concatenated to construct an overall feature map.To optimize the training time in MCNNE,we apply the principal components analysis(PCA)method for feature vector reduction.Combining the advantages of multiple sensors and the ensemble structure,MCNNE shows a great performance and eliminates the overfitting problem.Compared with single CNN structure and various ensemble bi-model structures,MCNNE has better performance.