Research And Implementation Of Fall Detection System Based On Fusion Decision

Falls and their subsequent problems are a serious threat to the health and life of the elderly.The research of falls detection technology can reduce the phenomenon of long-term lack of assis-tance after falls,make necessary medical help timely,reduce more possible harm and psychological shadow caused by falls,and improve the quality of life of the elderly.This paper investigates the latest developments in the field of falls detection at home and abroad,and chooses wearable devices as the technical direction of falls detection.Most of these studies are based on the changes of sensor data in the process of falls to detect falls.In these studies,a lot of work is based on customized sensors,which makes the system not universal and comparative;some work is based on intelligent devices which are more and more widely used in recent years,but there are still some problems such as low detection accuracy,strong intrusiveness of device placement location,unreasonable power consumption of the system and high reliability requirements of network transmission.In order to solve the above problems,this paper proposes a low-power threshold algorithm for loca-tion information aided decision-making and a classification algorithm based on GBDT-SVM fusion model,innovatively uses dual-track fusion decision-making mechanism,and realizes a fall detec-tion prototype system based on general intelligent equipment rather than special equipment.The main work and innovation of this paper are as follows:1)A method of extracting sequential signal window which is sensitive to falls and insensitive to daily activities is proposed to avoid the persistent waste of power caused by fixed time scale slid-ing window detection,which enhances the applicability of the algorithm on resource-constrained devices such as smart watches,reduces unnecessary detection and improves the detection accuracy to a certain extent.2)A threshold algorithm for location information aided decision-making is proposed.Aiming at the signal characteristics of falls,the characteristic indexes with better discrimination and less computation are selected.Taking the detection of falls and ensuring the life and health of users as the first criterion,a threshold algorithm with higher recall rate of falls is constructed.Then,by introducing location information,the human height information after falls is taken into account to exclude some non-falls and further fall.It reduces the false alarm rate of falls.3)A classification algorithm based on GBDT-SVM fusion model is proposed,which achieves very good classification results,and the model has been verified to have very good generalization performance for many kinds of falls and daily activities in reality.The GBDT model is used to au-tomatically combine the original features and map them to a new sparse feature space with higher dimensions,which enhances the ability of feature expression and improves the effect of classifica-tion and detection.At the same time，the sparse,"0/1" feature greatly accelerates the speed of model training and detection,and plays an important role in ensuring the timely response of information after remote processing when the system is implemented.4)In the system implementation,a dual-track fusion decision-making mechanism adapted to different scenarios and network conditions is proposed.On the one hand,low-power location information assistant threshold algorithm is implemented locally,on the other hand,more complex detection algorithm based on GBDT-SVM fusion model is implemented remotely when network conditions permit.The detection results of the two parts are independent,but the final decision fusion will be carried out according to certain rules,and the delay and unavailability of network connection are fully considered.Decide on the degree of intervention and reminder to enhance the overall availability and practicability of the system.In the actual scene verification of the prototype system,the detection accuracy of the fixed action type can reach 98%and the robustness of the unknown action type can also be guaranteed.