Short-time, Wavelet-inspired Mouse Submovement Detection
This work addresses a domain-specific challenge in motor interaction analysis, offering incremental improvements for researchers studying human motion in applications like first-person shooter aim data.
The paper tackled the problem of detecting overlapping submovements in one-dimensional speed time series, proposing a wavelet-inspired method with a self-weighted loss refinement step to improve accuracy, and demonstrated it on synthetic data with comparisons to existing techniques.
Submovements are ballistic components of human motion constituting a large part of motor interaction and arising from the cyclical and overlapping cognitive processes of perception, motor planning, and motor execution. Extracting submovements is challenging as the motions tend to overlap, or start before the previous ends. We propose and evaluate use of a wavelet-inspired technique to accurately locate and parameterize submovements from one-dimensional speed time series. Our method employs a self-weighted loss refinement step to identify and improve regions of poor quality of fit, a challenge for simpler wavelet transforms. We demonstrate the accuracy of our method by presenting analysis of ~6,400 1-2s trials of synthetic egocentric camera (first-person shooter) aim data for which we know ground truth, modeled from a similarly sized real data set of 13 users. We compare our method to dual-threshold and the persistence 1D segmentation techniques and note challenges and opportunities for future improvements.