Ambulation Assessment Using Deep Learning and Fiducials

On average, hospitalized geriatric patients spend over 85% of their time in bed. Frequent bed rest results in rapid physical deconditioning. To prevent this physical decline and the resulting complications, it is imperative that clinical staff set and monitor ambulatory goals with their patients. Often referred to as early and progressive ambulation, a plan to get patients out of bed and moving as soon as possible after injury or surgery greatly reduces the risk of additional complications and reduces hospital stays. Currently, no solution exists to effectively monitor and report ambulation in a clinical setting, greatly impacting clinicians’ ability to monitor patient progress. This work presents three systems for ambulation assessment based on three different technologies, real-time location systems, depth cameras, and RGB cameras. After examining the first two technologies a in-depth look into the implementation of RGB camera-bases ambulation assessment is taken. Leveraging a custom implementation of a deep neural network- based object detection algorithm facilitates detecting people and a set of assistive devices relevant to clinical environments. The object detections form the basis for the quantification of different ambulatory activities and related behaviors. Combined with simple fiducial markers and a Deep SORT algorithm, a subject can be detected and tracked simultaneously. Using features extracted from detected people and objects as input, machine learning models are leveraged to determine a person’s ambulatory status. This status includes motion, identity, distance traveled, posture, and use of an assisted device. The summary of an individual’s ambulatory status over time provides the data required to monitor overall ambulatory health and progress of hospitalized patients.