Automatic spline smoothing of non-stationary kinematic signals using bilayered partitioning and blending with correlation analysis

Measurement errors of kinematic signals introduced by motion capture systems are unacceptably amplified when differentiating the signal to derive acceleration. Existing fully automatic methods for solving this problem have a weakness on non-stationary signals involving impacts, while semi-automatic methods each require that users carefully determine the values of configurable parameters. We propose an automatic method that estimates acceleration from non-stationary kinematic signals using bilayered partitioning and blending (BPB) with correlation analysis. The method has a configurable parameter, the partition length, and the recommended value of the partition length that is empirically derived can be used without prior knowledge of kinematic signals. Having applied this algorithm to synthetic sinusoidal data, benchmarking data in the biomechanics community, and our own measurement data, we compared the results with those of existing automatic methods. The evaluation confirms that the proposed method estimates acceleration more accurately from non-stationary signals than existing fully automatic methods and is more robust than existing semi-automatic methods.