Silhouette classification using pixel and voxel features for improved elder monitoring in dynamic environments

We present a method for improving human segmentation results in calibrated, multi-view environments using features derived from both pixel (image) and voxel (volume) space. The main focus of this work is to develop a low-cost, vision-based system for passive activity monitoring of older adults in the home, to capture early signs of illness and functional decline and allow seniors to live independently. Silhouettes are extracted to address privacy concerns. Specific embedded assessment goals include daily gait, fall risk, and overall activity, as well as fall detection. To achieve these goals, accurate, robust segmentation of human subjects (silhouette extraction) from captured video data is required. We present a simple technique that makes use of features acquired from background subtraction results (silhouettes) of multiple calibrated cameras, along with the 3D voxel object formed from the intersection of those multiple silhouettes in a volume space to improve human segmentation results in dynamic environments; moving objects, non-human objects, and lighting changes often complicate this task. The technique is qualitatively evaluated on three data sequences, two of which were captured in an independent living facility for older adults.