PostureCheck: A vision-based compensatory-posture-detection tool to enhance performance of the BURT? upper-extremity stroke-therapy device - Summary / Abstract
This Small Business Innovation Research (SBIR) Phase-I project proposes the development of an image-
processing-based tool, named PostureCheck™, aimed at automatically detecting when patients perform
undesirable compensatory movements during robot-assisted upper-limb rehabilitation exercises. The system
will be based on a standard video camera (e.g., GoPro) that will be used to capture the movements of the
subject.
The automatic detection of undesirable compensatory movements is especially important when patients
use a rehabilitation robotic system with minimum supervision, i.e. when a single therapist oversees the
therapeutic sessions of multiple patients simultaneously. In this context, PostureCheck™ may be capable of
tracking robot-assisted rehabilitation exercises and enable feedback modalities to discourage the performance
of undesirable compensatory movements.
Our long-term goal is to integrate PostureCheck™ with the Barrett Upper-extremity Robotic Trainer -
BURT®, which we developed with special emphasis on stroke rehabilitation. The combination of
PostureCheck™ with the BURT® device would be ideally suited for deployment in “Robotic Gyms”, where a
single therapist oversees the therapeutic sessions of several patients simultaneously, thus allowing
rehabilitation centers to offer high-dosage, high-intensity interventions despite the limited number of therapists
currently available in the US.
To demonstrate the feasibility of the proposed concept, we will develop PostureCheck™ to detect the most
common compensatory movements automatically. To achieve this goal, we will rely on recently developed
artificial intelligence (AI) methods referred to as Deep Learning. These methods have recently broken records
in the human-posture analysis, joint-skeleton detection, and recognition of human activities using a single
inexpensive camera. The proposed video-based PostureCheck™ tool will be the first system to exploit the
capabilities of hybrid Deep Neural Networks, for real-time detection of compensatory movements during robot-
assisted rehabilitation.
The proposed SBIR Phase-I activities are organized in three aims. In Aim 1, feedback from rehabilitation
experts at Spaulding Rehabilitation Hospital will be used to collect video data and to label compensatory
movements observed during the performance of robot-assisted rehabilitation exercises by using the BURT®
system. In Aim 2, Deep Learning techniques will be used to develop a robust detection of undesirable
compensatory movements during the performance of robot-assisted rehabilitation exercises. Finally, in Aim 3,
the algorithms developed in Aim 2 will be optimized. Specifically, we will test implementations that are suitable
to generate real-time feedback. Computationally efficient implementations of the algorithms will enable - in
future studies - the development of new modalities of control of the rehabilitation robot with the objective of
discouraging undesirable compensatory movements.
