SVM Model for Feature Selection to Increase Accuracy and Reduce False Positive Rate in Falls Detection

Abstract

Falls are a dangerous problem for people of all ages. Thus, accurate falls detection with minimized false alarms is very important. This study aims to detect falls and activities of daily living (ADLs) using acceleration data and to introduce an effective feature selection criterion to reduce the false positive rate of the falls detection systems. The falls detection system in this study consists of three stages. At the first stage, we have harnessed some feature extraction techniques to have discriminative features from the acceleration data. Then we have used feature selection criterions to select effective features in the detection task. At the last stage, we used Support Vector Machine (SVM) to classify the selected features in falls and ADLs. We have used raw acceleration data and extracted all the features. Then we selected features based on the Minimum Redundancy Maximum Relevance (MRMR) criterion and Double Input Symmetrical Relevance (DISR) in the fall detection experiment. We have found that the DISR feature selection criterion is more effective in acceleration based fall detection system. The results show 100% classification accuracy and zero false positive rates in fall detection for the DISR based selected features.