Research On Fall Detection Methods Based On Wearable Computing

With the aging of China’s population and the increase in average life expectancy,the physical and mental health of the elderly is receiving attention from the government and all areas of society.In daily life,the occurrence of fall behavior is the most common and serious threat to the life safety and physical health of the elderly.Detecting,preventing,and intervening in real-time in the occurrence of falls can significantly reduce the risk of falls among the elderly and mitigate the harm caused by falls and the impact on the quality of life of the elderly.Wearable devices have been widely used in fall detection due to their portability,high privacy,and unrestricted detection area,but due to the slow walking of the elderly,the small range of motion of the arms,hips,and waist,and the insignificant differences in features between different behaviors,the following challenges still exist in the construction of models for fall detection of the elderly at this stage:(1)High false alarm rate of models.Some daily behaviors,such as running and climbing stairs,have high similarity with fall behaviors,so the models built using acceleration and other data can easily misclassify fall behaviors as non-fall behaviors or misclassify non-fall behaviors as fall behaviors during fall detection.(2)Poor generalization of the model.Numerous scholars have used subjects’ daily life behaviors and simulated or real fall behavior data to construct classification models to detect the occurrence of fall behaviors,but the models constructed using a certain training set cannot achieve the best detection effect on other users’ fall behaviors.Based on the above challenges,this paper segments different behaviors for the variability of behavioral features and achieves a more detailed segmentation of the fall process using the similarity of transient changes in acceleration data of different users’ fall behaviors to provide a new idea for fall detection for the elderly.The research of this paper is as follows:To address the challenge of the high false alarm rate of the model,this paper proposes a fall detection method based on dynamic sliding windows.When using fixed sliding windows to segment time series data,sliding windows of the same size cannot extract the best features of different continuous behaviors and cannot achieve the best classification effect for all behaviors,and the classification model can easily confuse fall behaviors with non-fall behaviors when the feature similarity between fall behaviors and certain daily behaviors is high.Therefore,in this paper,we construct sliding window sets based on the variability of the frequency domain features of different behaviors in the training phase and use multiple classifiers to train on different windows.In the testing phase,dynamic sliding windows are used to divide the temporal data,and the best window size for recognizing different behaviors is judged in realtime using the Fourier series,and the corresponding classifier is invoked for final recognition.Finally,this paper’s method is compared with the fixed-window and dynamic sliding-window methods on the Mobi Fall dataset,and the experimental results show that the accuracy of this paper’s method is improved by 6% compared with the fixed sliding-window method,and 5% and 3% compared with the other two dynamic sliding-window methods,respectively,and this paper’s research achieves a good recognition of fall-like behaviors in daily life This paper achieves good recognition of fall-like behaviors in daily life,thus effectively reducing the number of false alarms of fall behaviors.To address the challenge of poor model generalization,this paper proposes a fall detection method based on SAX-VAM.As a series of complex actions combined with abnormal behavior,falls have obvious user differences.In addition,there are many complex types of fall behaviors with different behavioral characteristics,which leads to difficulties in the detection of different users’ fall behaviors.Therefore,in this paper,we take advantage of the universal instantaneous acceleration variation of different users’ fall behaviors,introduce a peak detection process,group daily life behaviors,and fall behaviors using a finite state machine,transform time series into SAX words using a symbolic sequence representation and assign a binary label to the detected peaks using a SAX-VAM classifier,and finally perform an experimental validation on two publicly available datasets,TST and UMA Fall are experimentally validated and compared with two machine learning based fall detection methods.The experimental results show that the optimal classification accuracy of the method in this paper is 97.05%on the TST dataset and 95.79% on the UMA Fall dataset for fall behaviors.The excellent performance on the above different datasets reflects the better generalization of the classification model.