A Wearable Gait Monitoring System Based on Computer Vision

Introduction: Biomechanical features of human gait are commonly used to assess musculoskeletal and neurological function, disease progression, and age related biological factors. Gait disorders are commonly experienced by older adults, increasing their risk of falls and fractures, resulting in a lower quality of life, limited ability to conduct daily activities, and increased morbidity. The ability to monitor and detect gait disorders over different phases of walking plays a crucial role in the diagnosis and treatment of disease. Current gait monitoring technologies have substantial drawbacks in hardware complexity and functionalities. For example, commercial inertial measurement units (IMUs) experience drift and difficulty directly correlating signal features to known gait characteristics or events. Gait mats have fixed length and are limited to a few steps of measurement. Optical motion capture systems generally require a highly controlled environment. Most of these systems are also prohibitively costly.

To counter these drawbacks, we demonstrate a new stereo-camera based, shoe-mounted wearable gait monitoring system that can accurately measure up to 15 gait parameters. Data from the system are highly reproducible and demonstrates low signal drift. It can handle walks longer than the typical gait mats, allowing for capture of a richer gait data set, which can be incorporated into current large language models to facilitate the diagnosis of gait diseases. The system is low-cost (less than $200), easy to use, and suitable for real-life measurement in both the hospital and home settings.

Materials and

Methods: The system has two wearable sensor systems, one for each shoe, as shown in Fig. 1. Each wearable has a stereo-camera system consisting of a pair of Raspberry Pi Zero 2 W, large-field-of-view (FoV) IR cameras, and 1200mAh PiSugar3 portable batteries. The 120°-FoV cameras are triggered by a polymer force sensing resistor (FSR) placed at the heel. The FSR circuit detects the force level applied to the FSR, records the time stamps, and determines the onset of a gait event, resulting in the camera on one foot taking images within 0.1 second of a marker placed on the other foot. The system uses hardware interrupts, universal asynchronous receiver-transmitter (UART) communication protocol, and user datagram protocol (UDP) to interrupt and transmit data.

The markers were designed specifically to optimize the real-time detection and measurement of the spatial gait parameters. We collected 1,600 images of the markers from different participants walking while wearing the system. The images and augmented images were fed into the YOLO deep learning model to train for real-time object detection. The pixel positions of the markers from the object detection (from both cameras) were combined to compute the 3-axis marker positions and corresponding gait parameters through the stereo camera algorithm. Spatial parameters were combined with the time stamp data for the extraction of temporal gait parameters such as gait speed. Parallel computing that assigns the gait parameter calculation to both Raspberry Pis was applied to optimize the computation loads.

Results, Conclusions, and Discussions: Results/Discussions
Six healthy volunteers participated in two experiments to demonstrate the gait monitoring system. As shown in Fig. 2. (a), the first experiment explores the accuracy of 15 measured gait parameters. Participants wearing the gait monitoring system were asked to walk back and forth eight times on a standard gait mat. The gait parameters measured from our system were compared with those values acquired from the gait mat to estimate the accuracy of our system. The results of the average accuracy are shown in Fig. 3.

The second experiment (Fig. 2. (b)) investigates the repeatability and drift of the system over long walks that began and ended on a gait mat. In the middle, the participants walked off the gait mat for around 120 steps and made 3 turns. We compared the accuracy of our system at the beginning and the end of the long walks to validate the consistency of the system. The results showed, on average, a 1.48% drift across the 15 parameters between the beginning and the end of the test. This demonstrates that our system can be used for long walks to collect accurate gait data for long gait sequences, with minimal drift.

The deviations in the calculated results from the ground truth data may be attributed to a couple of factors. The stereo camera might have shifted slightly from the original place when participants were walking, introducing errors to the spatial parameters. The placement of the FSR might have changed because of the different types of shoes of participants, resulting in errors in timestamp capturing. In future work, we plan to solve these problems, test our system in people who have gait disorders, collect long-sequence gait data, and design and train transformer models for gait disease diagnoses.

Conclusions
In conclusion, we have demonstrated a new type of stereo camera gait monitoring system based on computer vision. The system can measure 15 gait parameters, all with an accuracy greater than 91%, high repeatability and low drift for long walks. The proposed gait monitoring system is low cost, easy to use, and suitable for real-life measurement.

Acknowledgements (Optional):