| Among Elderly population,falling is a common and dangerous event that will cause great damage to their bodies.However,in modern society,most young family members will not stay at home all the time because of work or study tasks,and the elder people spend more time alone at home.When the elder people fall down and get hurt,they will have serious health problems if they can’t get help in time.The risk can be effectively reduced by detecting the fall of the elder people and sending out a distress signal in time through technical means.Using wearable sensor information to detect falls is one of the mainstream methods at present.In this thesis,based on the acceleration information and angular velocity information of the waist,a feature construction method is proposed,and the model is established by using LightGBM classification algorithm.It is implemented on the public data set FallAllD and compared with other methods to verify the effectiveness of the proposed method.First,the original data frequency is processed into 200 HZ,100HZ,50 HZ,40HZ,25 HZ and 20 HZ,and the features are extracted and screened respectively,and the filtered features are used for LightGBM modeling.According to the Averecall index,the model obtained from the data with frequency of 200 HZ has the highest value on the verification set,reaching 97.63%.Then,based on the previous six models,the voting method integration models under different combinations and voting thresholds are studied.The averecall index reaches the best result in the integrated model with voting threshold of 2 consisting of 200 HZ,100HZ,50 HZ,40HZ and 20 HZ,up to 98.34%,which is 0.71% higher than the single model.Single models with different frequencies are at least 1.14% higher than the benchmark models,and the best results from the integration methods are 2.11% and 3.17% higher than those from the two benchmark models,respectively. |
