The evaluation of depth image features for awakening event detection

Abstract

Falls among bedridden would increase in number if they are left unsupervised by the caregivers. The aim of this study is to evaluate the features from the Kinect-like depth image representing the bedridden in detecting the awakening event as the event that falls might occur. The images from 20 subjects performing six sleeping activities including the awakening events were obtained before image segmentation based on horizontal line profile was computed to these images in localizing the bedridden as region of interest. After that, the biggest blob selection was executed in selecting the biggest blob (blob of bedridden person body). Finally, blob analysis was formulated to the resultant image before boxplot and machine learning approach called decision tree were used to analyze the output features of blob analysis. Based on the results from the boxplot analysis, it seems that centroid-x is the most dominant feature to recognize awakening event successfully as the boxplot represent the centroid-x of awakening event were not overlap with other sleeping activities. The result from machine learning approach is also seem in good agreement with boxplot analysis whereby the modelled decision tree with solely using centroid-x achieve the accuracy of 100%. The second largest accuracy is the perimeter followed by major axis length and area.