A Comparison of Smoothing and Filtering Approaches Using Simulated Kinematic Data of Human Movements

Gathered kinematic data usually requires post-processing in order to handle noise. There a three different approaches frequently used: local regression & moving average algorithms, and Butterworth filters. In order to examine the most appropriate post-processing approach and its optimal settings to human upper limb movements, we examined how far the approaches were able to reproduce a simulated movement signal with overlaid noise. We overlaid a simulated movement signal (movement amplitude 80 cm) with normal distributed noise (standard deviation of 0.5 cm). The resulting signal was post-processed with local regression and moving average algorithms as well as Butterworth filters with different settings (spans/orders). The deviation from the original simulated signal in four kinematic parameters (path length, maximum velocity, relative activity, and spectral arc length) was calculated and checked for a minimum. The unprocessed noisy signal showed absolute mean deviations of 54.78% ± 12.16% in the four kinematic parameters. The local regression algorithm revealed the best performance at a span of 420 ms with an absolute mean deviation of 2.00% ± 0.86%. For spans between 280–690 ms the local regression algorithm still revealed deviations below 5%. Based on our results we suggest a local regression algorithm with a span of 420 ms for smoothing noisy kinematic data in upper limb performance, e.g., activities of daily living. This suggestion applies to kinematic data of human movements.