ECSS Paris 2023: CP-BM12
INTRODUCTION: Working with heterogeneous sensor data can substantially improve movement analysis by providing richer biomechanical detail[1]. A persistent obstacle in such setups is temporal misalignment between data streams[2]. Even small offsets can affect outcomes and worsen biomechanical predictions. Current approaches typically (i) rely on explicit synchronization actions or identifiable shared events to align sensor signals[2] or (ii) use hardware-based synchronization via a master device and wireless protocols[1]. However, controlled synchronization actions are often impractical in field or laboratory settings due to missing or mistimed actions, while hardware solutions are not consistently available across systems. Consequently, reliable post-hoc synchronization remains crucial for accurate analyses. METHODS: In this work, we adapt the Nearest Advocate algorithm (NAd)[3], an event-based time delay estimation method originally developed for multi-sensor event time series, to sport science. Raw pressure measurements were preprocessed using a Butterworth low-pass filter, followed by a temporal differentiation, half-wave rectification, and moving-average smoothing to enhance heel-strike features. Heel-strike events were then identified via peak detection with the minimum peak distance constrained by a maximum expected step frequency of 300 spm. Experiments were conducted using 40 instrumented treadmill pressure signals as reference, recorded across multiple speeds and incline conditions per run. For controlled evaluation with known time delays, clones of these signals were generated that are temporally shifted and noise-contaminated, with additive noise levels up to 10% of peak pressure. The experiments comprised: (i) estimation of randomized time delays, (ii) estimation of randomized time delays under systematically missing steps (removal of every second step to simulate one-sided pressure signals), and (iii) visual and statistical assessment of alignment between these treadmill recordings and corresponding pressure insole (PI) data. RESULTS: NAd accurately recovered (i) randomized time delays (RMSE = 0.81 ms ± 0.80 ms; MAE = 0.16 ms), with (ii) moderate degradation when every second step was removed (RMSE = 1.79 ms ± 1.78 ms; MAE = 0.67 ms). For real data, (iii) the temporally corrected (one-sided) PI-signals aligned clearly with treadmill recordings, yielding low post-alignment residual errors (RMSE = 0.24 ms), while estimated global device offsets were ~0.75 s (right) and ~0.74 s (left) with substantial between-trial variability (SD = 1.14 s). CONCLUSION: These experiments indicate that NAd enables millisecond-level post-hoc time delay correction of heterogeneous sensor signals, without requiring dedicated synchronization actions or additional hardware. This enables practical synchronization in cross-device measurement setups and supports more reliable downstream data analysis. [1] Verdel 2023 [2] Bannach 2009 [3] Schranz 2024
Read CV Christina HalmichECSS Paris 2023: CP-BM12