A comparative study on human motion smoothing filters for RGB-D camera teleoperation

RGB-D camera teleoperation is a promising marker-free method to control robots by capturing human motion. However, the motion-captured data tends to oscillate significantly, leading to a significant deviation in a single step, which cannot be directly used to control the robot in Cartesian space in real time. This paper comprehensively evaluated time-series data filter algorithms to alleviate the data oscillation issue. We compared eight mainstream filters, namely the Savitzky-Golay (S-G) filter, hierarchical threshold filter, median filter, Whittaker filter, weighted sliding convolutional averaging (WSCA) filter, adaptive filter, Kalman filter, and locally estimated scatterplot smoothing filter, with four performance merits (time complexity, mean absolute error (MAE), variance (VAR), and root-mean-square error (RMSE)). Here, the YOLOv8-pose human skeleton recognition algorithm was employed to test different filters due to the excellent motion-capture performance. We conducted 12 groups of human-motion experiments with circular trajectories. The results indicated that the S-G filter excelled in data smoothing with a 26.74% improvement, followed by the WSCA filter with a 24.4% improvement. Meanwhile, the time complexity of the S-G and WSCA algorithms was lower than that of the other six filters. Furthermore, we tested these two algorithms in handwritten teleoperation. The results demonstrated that the filtered trajectory was smooth for the handwritten tasks and verified the effectiveness of the SG and WSCA filter algorithms.