Research On Fall Detection Based On Long Short-term Memory Artificial Neural Network And Wrist Sensor

The aging trend of population in our nation is accelerating.Falls pose a significant threat to the health of the elderly.Lying on the ground for a long time after a fall will cause further harm,which migth even lead death.Therefore,the research on fall detection is necessary.Wearable sensors such as Inertial Measurement Unit(IMU)have been widely used for fall detextion.It is difficult to balance the false alarm and miss detection of falls.Besides,most studyies used trunk and back as the position to put the sensors.With the popularity of smart wristband watches,it is undoubtedly of high application value to detect falls through the inertial sensor on the wrist,and by adopting the deep learning algorithm,high-level features can be automatically obtained from the data during the selection of fall detection indicators.Therefore,the objective of this research is to develop a fall detection model based on IMU on the wrist and Long-Short Term Memory model and verify the validity of the model.Specifically,the contents of this research include:(1)design the experiment,fifty young adults and twenty older adults were recruited.Young participants were asked to take unintentional and intertional falls,and daily movement,IMUs were put on their wrsit to collect the inertial sensor data for model development.(2)The IMU data were pre-processed.Eventually,it resulted in a total of 398 data samples which had unified data length.Specifically,the data contained 55 slips and fall,57 trip and fall,21 falling forward,25 falling left,23 falling right,20 falling backward and 16 falling slowly against the wall,20 going down stairs,19 going up stairs,18 sitting down quickly,19 sitting down and standing up,20 crouching down and standing up,16 lying down,20 hitting the mat,53 jumping,20 two-handed applause and 20 times of two-handed wave movement.(3)A fall detection model was deleloped based on LSTM and evaluated test set and verification set.The results showed that the precision rate of the model for fall judgment was 85.72%,the accuracy rate was 85.90%,the recall rate was 85.90%,and the F1 value was 85.79%.The main innovations of this study are as followed:(1)In the experimental design,passive fall and active fall are introduced at the same time to make the scene of fall closer to the real fall.By introducing more real fall data,the generalization ability of the model is guaranteed.And 20 elderly subjects were recruited to obtain the data of daily movements of the elderly,which ensured the diversity of experimental data samples.(2)Based on potential application scenarios such as smart bracelet and watch,the data collected by wrist sensor was selected as data samples,and a fall detection model was established based on deep learning network.Compared with the existing fall detection research,this study has data and the innovation of algorithm.In summary,abundant fall data and daily action data were collected in this research including the unitentional falls,which was more similar to the real fall.Wrist was used as the IMU attachment postion which was more friendly and with high application value.The fall detection model was based on LSTM,a more advanced deep learning model is selected.And multiple indicators are used to evaluate and analyze the model.Finally,it is verified that the fall detection model proposed in this study can accurately judge whether a fall occurs according to the data rules generated in different scenarios.