Research On Human Fall Detectionbased On Floor Vibration

With the continuous development of economy and society,the aging of population and the phenomenon of "empty nest" are becoming increasingly obvious.The health and safety of the elderly have aroused widespread concern in society.Falls are one of the main risks threatening the lives of the elderly.Falls not only cause physical damage to the elderly,but also threaten their lives because they are not treated in time.Fall identification method,on the one hand,can let the fallen elderly get the first time treatment,improve the level of fall relief,reduce the possibility of greater injury caused by falls,while reducing the occupancy of medical resources,reducing their family burden;on the other hand,can provide security for the elderly,especially the empty nest elderly.It helps to lighten their psychological burden and improve their quality of life in their later years.According to the different channels of signal acquisition,the existing fall recognition methods are mainly divided into three categories:(1)fall recognition based on video images;(2)fall recognition based on wearable equipment;(3)fall recognition based on floor vibration.The problem of first two methods are not technical problems,but the fatal defect of revealing privacy or forgetting to wear.Therefore,this paper studies the problem of fall recognition from the perspective of floor vibration.Firstly,the fall load model is established,and the database of fall recognition is established by numerical simulation.According to the database,the feature selection and algorithm design are carried out.The peak value,energy and sensor correlation coefficients are selected as features by the method of inter-class divergence,and the multi-feature set semi-supervised SVM algorithm is proposed.Finally,the experimental database is used to verify the algorithm.Previous scholars have proposed benchmark problem based on floor vibration,and established a benchmark database including ball drop,bag drop and human free jump.However,the benchmark database does not contain human fall behavior,so the proposed recognition algorithm lacks pertinence.In this paper,a semi-supervised SVM fall recognition algorithm based on multi-feature set is proposed,which makes up for the base database by numerical simulation of human fall behavior.The advantages of this method are verified by comparing with the benchmark problem algorithm.Finally,the method is further validated by experimental data.Aiming at the problem that the sponge cushion is used to weaken the vibration signal of the floor in the human fall experiment,a falling load model is proposed in this paper.Based on the qualitative analysis of the fall process,the fall load modelis proposed and verified by the fall database based on wearable equipment.Aiming at the problem of benchmark problem with the background of fall recognition but not involving the fall experiment,the ANSYS finite element model of the floor of benchmark problem laboratory is established to simulate the floor vibration under three kinds of human-induced loads: fall,walk and rhythmic jump,and the acceleration time history data of sensor nodes in the benchmark problem are extracted.With the three activities of packet dropping,ball dropping and free jumping in the benchmark problem database,the database of fall recognition based on benchmark problem is composed.Concerning about the database of fall recognition based on benchmark problem,peak value,energy and sensor correlation coefficient are selected as classification features by the method of inter-class divergence,and a multi-feature set semi-supervised SVM algorithm is proposed.Aiming at the problem that the SVM algorithm is sensitive to the number of training samples but can not obtain a large number of training samples in reality,the semi-supervised learning strategy is adopted;the strategy of constructing base classifier based on multi-feature sets is adopted for the different false judgment rate of non-falling events in the recognition results of SVM algorithm based on a single feature set.The results show that the proposed multi-feature set semi-supervised SVM algorithm has a good learning ability,and the recognition accuracy is significantly improved under different markup rates,with an average increase of 9.68%.At the same time,compared with the benchmark problem algorithm,it has obvious advantages in the sensitivity of fall recognition.Because the algorithm is based on the database of benchmark problem,and the fall,walk and rhythm jump in the database are obtained by numerical simulation,the algorithm is further verified by the experimental database.The results show that the algorithm proposed in this paper can effectively improve the recognition accuracy under different marking rates,and has good noise resistance,and the false alarm rate is significantly higher than the false alarm rate.Dropping of bags and balls are the main factors that cause false positives in non-falling events.