Pre-Impact Fall Detection Based on a Modified Zero Moment Point Criterion Using Data From Kinect Sensors

Accidental falls have always been a serious problem for the elderly. There is considerable demand for pre-impact fall detection systems with long lead times. According to the zero moment point criterion, the zero moment point should be kept beneath the supporting foot for stability during humanoid robot standing or walking. However, the zero moment point in the human walk does not stay fixed under the supporting foot. In this paper, we define a dynamic supporting area containing both feet and the area between the two feet, and propose a method of fall prediction based on a modified zero moment point criterion using motion-monitoring data from a Kinect sensor. A fall event is predicted if the projection of the zero moment point locates outside of the dynamic supporting area. The proposed method is compared with a method identifying the imbalance state based on a support vector machine classifier. Experimental results show that fall events could be detected with an average lead time of 867.9 ms (SD = 199.2), a sensitivity of 100%, a specificity of 81.3%, a positive predictive value of 87.0%, a negative predictive value of 100%, and an accuracy of 91.7% using the modified zero moment point criterion. The lead time was 571.9 ms (SD = 153.5) and accuracy was 100% for the support vector machine classifier. The modified zero moment point criterion-based method achieved the longest lead time in the pre-impact fall detection.