StrollerPulse

StrollerPulse is an open-source, audio-first terrain vibration monitoring tool that uses the accelerometer and gyroscope built into ordinary smartphones to monitor terrain vibration and announce it by voice.

StrollerPulse uses Prior Labs' TabPFN tabular foundation model for real-time terrain classification of sensor features.

StrollerPulse computes a 6-dimensional feature vector from a 2-second sliding window (100 samples at 50 Hz), strictly following the ISO 2631 human vibration standard, and classifies three terrains: natural grass trail, smooth asphalt pavement, and severe cobblestone vibration hazard zone.

StrollerPulse uses a two-stage step-frequency decoupling filter to reduce the false alarm rate caused by human step frequency from 38.4% to 0.0%, an improvement of 18.3 times.

The two-stage step-frequency decoupling filter separates human step-frequency harmonics through low-band (1.5–3.5 Hz) moving spectral autocorrelation, then uses scipy.signal.filtfilt to apply a fourth-order zero-phase high-pass Butterworth filter (cutoff frequency 6.0 Hz) to attenuate human swaying while preserving mechanical shock signals in the 12–28 Hz band.

In a benchmark of 30 held-out test windows, TabPFN achieved 100.0% overall accuracy, 100.0% shock recall, 0.33 ms CPU latency, 0.0% step-frequency false alarms, and required no training or tuning.

In a 1.4-kilometer community walk field test, StrollerPulse's G-Shock severity held steady at 22/100 during the concrete sidewalk phase, and when stepping onto the grass shoulder the variance dropped sharply from 0.048 to 0.012, with the headset announcing within 1.5 seconds that frame vibration had dropped 44%.

StrollerPulse's speech synthesizer initially announced every 2 seconds on grass, and was later changed to event-driven debouncing, triggering speech only on terrain state transitions or severe impact alerts.