Research And Implementation Of Fall Detection Algorithms

The increasing number of aging people has been a serious problem for Chinese society.According to research,fall is one of the main threatens to elders' health.However,due to the complex and random features of fall behaviors,it is hard to detect fall accurately with traditional fall detecting algorithms.To improve the detection accuracy,we propose a new fall detection algorithm based on acceleration.In this paper we analyze various features of fall behaviors and choose acceleration as the most significant indicator.We extract the time domain and waveform features of fall behaviors' acceleration data.The extracted features are selected by a selecting algorithm.This algorithm guarantees minimum redundancy and maximum correlation of outputs through choosing maximal mutual information and information entropy.Treating the selected features as training data,we test some classic pattern recognition algorithms to categorize fall and non-fall behaviors.The test results show that SVM and Adaboost perform better than others.Concerning that fall detection involves unbalanced data and unequal mis-classification cost,we modify the traditional SVM and Adaboost algorithms to propose Cost Sensitive SVM(CS-SVM)and Cost Sensitive Adaboost(CS-Adaboost)algorithms respectively.The CS-SVM takes advantages of cost function and minimum risk Bayes decision to improve SVM,while the CS-Adaboost enhances Adaboost through involving cost function and unbalanced strategies.The test results show that both modified algorithms categorize better than original algorithms and the CS-SVM performs best.