Research Of Feature Set Optimization And Implementation Technology Of Falling Detection Based On Mobile Phone

Real-time detection of the elderly fall behavior and informing the relevant personnel for treatment in time after the fall occurred is of great significance of the health care of the elderly.Traditional fall detection methods due to the need to add additional testing equipment,there are some problems in the cost of equipment or the feasibility of the system.As the proliferation of smart phones,the growing computing power and the diversity of the means of communication and built-in sensor,more and more scholars focus the equipment of fall detection on smart phones.Limited by the calculation of mobile phones and endurance,fall detection algorithm must be designed to reduce the power consumption of mobile phones on the basis of ensuring good classification result.Therefore,in this thesis,a fall detection combining threshold judgment and pattern recognition is designed.Classification model is one of the most important factors influencing the detection accuracy and computational complexity.In order to obtain the classification model with simple calculation and good classification result,it is necessary to improve the characteristic quality of the input sample.In this thesis,a 27-dimensional feature set is extracted from the time domain and frequency domain in combination with literature and a large number of experimental analyzes.In order to eliminate the redundant and non-relevant features which are concentrated in the feature set,this paper proposed an optimized feature selection algorithm based on neighborhood granulation and discrete binary particle swarm optimization(BPSO),and using the algorithm,the most representative feature subset of six dimensions has been obtained.In the system implementation,this paper also optimized the power consumption of training data acquisition,behavior data segmentation and real-time fall detection.A fall detection system based on the android platform has been designed and realized.According to test results,it comes with an average of 98.7%sensitivity and 98.5%specificity.Specifically,this paper mainly involves the following work:1.A feature selection algorithm based on neighborhood granulation and BPSO is designed,and the heuristic forward search function algorithm is used to optimize the problem that BPSO is easy to converge to local optimal solution and the result is inconstant.2.A three-level fall detection algorithm combining threshold judgment,pattern recognition and post-fall correction is designed and implemented,which not only reduces the power consumption of data processing,but also ensures the correct rate of fall detection.3.A platform combined with multi-behavior data acquisition and model training is set up.With this platform,the data collection,data storage,data view,data analysis and model training can be completed.4.The fall detection system based on Android smartphone with real-time fall detection,alarm,all-day fall guard and model online learning is designed and realized,and the power consumption of the related process is optimized.