| As the common unintentional injury,falls are one of the main causes of bruises and fracture among the elderly.Fall detection system can detect the occurrence of fall events immediately,so as to reduce the injury of the elderly.In practical application,wearable fall detection system is faced with many problems,such as low detection accuracy,high false alarm rate and high power consumption.Therefore,a fall detection method is proposed in this paper and a fall detection system combining wearable devices(placed at the waist)and cloud platform is deployed.To solve the problems of slow processing speed and short battery life,a low-power wearable device and its matching threshold based method(TBM)is designed.The low-power technologies based on hardware are employed,including the selection of low-power electronic devices,the use of hardware interrupt technology instead of software program real-time monitoring,and prolonging the sleep time of the data transmission module.For TBM,both the power consumption of the feature extraction method and classification accuracy are considered,and the optimal fall-related characteristics are chosen.Furthermore,the particle swarm optimization algorithm is applied to optimize the thresholds of TBM.As a result,the activated number of wireless transmission is reduced,while almost all fall event is uploaded to the cloud platform.Experimental results show that the TBM deployed in wearable device successfully identified 99.64% of fall events and rejected 82.47% of activities of daily life(ADL).To improve the accuracy of fall detection technology,two machine learning based fall detection algorithms were proposed: the fall detection algorithm based on recurrent neural network(RNN)and the fall detection algorithm based on transfer learning(TFL).In the offline test,the sensitivity and false alarm rate of TBM+RNN were 98.26% and 0.83%,respectively.The sensitivity and false alarm rate of TBM+TFL were 95.83% and 1.12%,respectively.In addition,RNN was deployed on the cloud platform and the fall detection system(wearable device + cloud platform)was tested online.Experimental results show that the sensitivity,false alarm rate and accuracy of the system are 100%,0.83% and 99.50%,respectively. |
