Download Full Text (191 KB)

Faculty Advisor(s)

Kurt Jackson, PT, PhD


Purpose/hypothesis: The ability to monitor human activity remotely may be useful in providing telerehabilitation and measuring real-world rehabilitation outcomes. Current methods of activity monitoring have significant limitations (cost, privacy, ease of use) that can limit their benefit and widespread use. Recent advances in machine learning (ML) and millimeter wave radar (MWR) have allowed for the development of a cost effective and simple way to monitor human movement continuously and remotely while maintaining reasonable privacy. The purpose of this study was to test the ability of ML and MWR to accurately classify and monitor different types of human activity including different gait patterns to monitor for changes over time.

Number of Subjects: 74 healthy adults (mean age = 24 ± 7.36, range = 21-53)

Methods and Materials: A skeleton pose estimation (Microsoft Kinect) and micro-Doppler signatures were combined to train the program to recognize and accurately classify five different gait patterns. The five gait patterns included a normal gait, a limping (antalgic) gait, a stooped posture, with the use of a walker, and with the use of a cane. Subjects spent one minute in each of the following scenarios: walking perpendicularly, parallel and freely in front of the sensor. This design allowed data to be collected from multiple angles to improve the model’s accuracy in all planes.

Results: A real-time simulation was performed to observe the point cloud behavior for different activities which then validated the system against the ground-truth values. Therefore, this allowed the system to calculate the training and prediction accuracy levels. Accuracy of the system was determined for each gait pattern using ROC/AUC analysis and were as follows: Normal = 97.2%, Stooped = 95.7%, Limp = 97.4%, Walker = 98.8%, and Cane = 98.4%. The first three activities (normal, limping, and stopped gait) demonstrated slightly lower prediction accuracy levels due to the gait patterns overlapping.

Conclusions: Prediction accuracy values ranging from 95.7 - 98.8% demonstrated good to excellent ability of the millimeter wave radar and machine learning model to classify different gait patterns in an open environment. Additionally, it has the capability to track the gait patterns observed and the changes that occur in a space simultaneously.

Clinical Relevance: An inexpensive radar system could be used in both home and institutional and community settings to accurately monitor activity levels and detect changes in gait over time. The ability to detect changes in gait patterns over time could be helpful in recognizing declines in function that require intervention or for monitoring outcomes of rehabilitation.

Publication Date



Physical Therapy | Rehabilitation and Therapy

The Use and Accuracy of Millimeter Wave Radar and Machine Learning for Gait Classification and Monitoring