Noise Reduction in Human Motion-Captured Signals for Computer Animation based on B-Spline Filtering

Mehdi Memar Ardestani, Hong YanSensors 2022 · DOI

TL;DR

This paper tackles the pervasive issue of noise in human motion capture data, which hinders its use in computer animation and virtual reality. The authors propose and evaluate B-spline filtering as a superior method for smoothing this data compared to traditional moving average and Kalman filters, offering better control over smoothness and continuity.

What Problem Does This Paper Address?

Motion capture (MoCap) is crucial for creating realistic animations and immersive virtual reality experiences. However, raw MoCap data is often corrupted by noise introduced during the capture process. This noise can stem from various sources, including sensor inaccuracies, environmental factors, and subject movement. Before this data can be effectively used, it needs to be processed to remove these artifacts. Existing methods like moving average and Kalman filters have limitations, such as introducing phase shifts or struggling with sudden signal spikes. The paper aims to find a robust and flexible filtering technique that can produce high-quality, smooth motion signals suitable for downstream applications like derivative calculation for animation.

Methods

The paper compares three primary filtering techniques:

  • Moving Average Filter: A simple method that averages data points within a sliding window. The authors explore both symmetric and asymmetric windows of varying sizes.
  • B-Spline Smoothing: This method approximates the signal using piecewise polynomial curves defined by control points. The key advantage highlighted is the ability to control the degree of continuity and flexibility of the resulting curve. The authors use a least-squares approach to find optimal control points.
  • Kalman Filter: An iterative process that estimates the state of a system from noisy measurements. It's known for its ability to provide online estimates.

The performance of these filters is evaluated using two metrics: SM1 (measuring signal smoothness by summing absolute differences between consecutive points) and SM2 (measuring signal energy, related to curvature). Percentage differences (%SM1 and %SM2) are used to quantify the improvement achieved by filtering.

Datasets

The paper uses two types of datasets for evaluation:

  • Simulated Noisy Sine Signal: A synthetic dataset created by adding random noise to a clean sine wave. This serves as a controlled environment to test the fundamental filtering capabilities.
  • Real-World Motion Capture Data: Data from the Carnegie Mellon Motion Capture Database, specifically a "kicking soccer ball" sequence. This dataset captures complex human motion with multiple degrees of freedom, providing a more realistic test case.

Key Insights and Ideas

  • B-spline's Flexibility is Key: The B-spline method offers superior control over the smoothness and continuity of the filtered signal, which is crucial for applications requiring accurate derivative calculations (e.g., velocity, acceleration).
  • Moving Average Limitations: While computationally inexpensive, the moving average filter struggles to effectively remove high-frequency noise, especially in complex motion sequences.
  • Kalman Filter's Trade-offs: The Kalman filter can provide online estimates, but its performance is sensitive to the measurement error parameter, and it can introduce lags in the output.
  • Parameter Tuning is Crucial: The effectiveness of all filters is highly dependent on their specific parameters (window size for moving average, number of control points for B-spline, measurement error for Kalman). Finding the right balance is essential to avoid over-smoothing or under-smoothing.
  • B-spline for Derivative Calculation: The paper emphasizes that B-spline smoothing is particularly advantageous when accurate derivatives of the motion signal are needed for downstream tasks like clustering or classifier training.

Key Results

  • B-spline Outperforms: In both simulated and real-world scenarios, B-spline smoothing generally achieved better smoothness metrics (lower SM1 and SM2 values) compared to moving average and Kalman filters, especially for complex motions.
  • Moving Average Ineffective for High Frequencies: For the real motion data, the moving average filter showed limited effectiveness in removing high-frequency movements, even with larger window sizes.
  • Kalman Filter's Sensitivity: The Kalman filter's performance varied significantly with the Emea parameter. While it could reduce noise, it sometimes introduced lags and occasional noisy movements.
  • Optimal B-spline Parameters: For the B-spline method, using a smaller number of control points led to smoother curves, but care must be taken to avoid over-smoothing that deviates from the original signal's trend. The paper suggests that a few numerical experiments are needed to find the optimal number of control points.

Limitations and Future Work

  • Offline vs. Online: The B-spline method, as implemented, is an offline process, meaning it requires the entire signal to be available before processing. This limits its use in real-time applications where online filtering is preferred.
  • Parameter Selection: While the paper demonstrates the impact of parameter tuning, it doesn't provide a definitive automated method for selecting optimal parameters for B-spline smoothing across all motion types. Future work could focus on developing such automated parameter selection strategies.
  • Generalizability: The study focuses on specific types of motion (walking, kicking). Further research could explore the performance of these filters on a wider range of human activities and motion capture setups.

Summary updated

BibTeX
@article{ardestani2022noise,
  author = {Mehdi Memar Ardestani and Hong Yan},
  title = {{Noise Reduction in Human Motion-Captured Signals for Computer Animation based on B-Spline Filtering}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22124629},
}
Figure