Objective: This study investigates the effects of various filter types on data obtained from a webcam-based 3D gait analysis system.
Background: Traditional motion capture systems, such as those utilizing infrared optical motion capture cameras or IMU sensors, are not only costly but also invasive.
By employing a standard RGB camera with OpenPose, an open-source system for human body segment detection, it is possible to explore alternative options for professional motion capture systems. However, the resulting data often exhibit significant noise, necessitating the application of suitable filters and parameters.
Method: Body segment coordinates were acquired during two complete gait cycles while subjects walked at a speed of 5km/h. The data were subsequently filtered using four distinct filter types: Butterworth, Gaussian, LOESS, and Median filters, each with varying parameters. The processed data were utilized for inverse kinematics calculations to determine the angles of the hip, knee, and ankle joints.
Results: The application of a 2nd or 4th order low-pass Butterworth filter, with cutoff frequencies ranging from 2 to 15Hz, a Gaussian filter with sigma values between 5 and 10, and a LOESS filter with window sizes of 5 to 20 resulted in smooth movements. Furthermore, the low-pass Butterworth filter, with cutoff frequencies between 5 and 15Hz, exhibited the most seamless motion and yielded low RMSE values. While the Median filter with a window size of 9 produced the lowest overall RMSE value, it failed to effectively remove continuous noise throughout the data, resulting in unnatural movements.
Conclusion: When filtering gait data obtained from a webcam-based system, utilizing filter parameters within the suggested ranges for each filter type can promote both fluent motion and accurate angle values.
Application: The findings of this study provide threshold guidelines for each filter type within a webcam-based gait analysis system.