aliro_gait.py

Cclass WalkUp

One analyzed walk-up: identity, series shape, features (or skip reason).

F_solve3(a, b)

Gaussian elimination for the 3x3 normal equations of the quadratic fit.

F_polyfit2(ts, xs)

Least-squares x(t) = a*t^2 + b*t + c; returns (a, b, c).

F_resample(ts, xs, dt)

Linear interpolation onto a uniform dt grid (missed blocks leave gaps in the capture; the spectrum needs even spacing).

F_dft_power(xs, nfft)

Power spectrum, bins 0..nfft/2 (input zero-padded to nfft). O(N*K) is fine at these sizes; keeps the tool stdlib-only.

F_norm_spectrum(xs, nfft)

Hann-windowed magnitude spectrum normalized by coherent gain, so a unit-amplitude sine at a bin center reads ~1.0 regardless of length.

F_diff_spectrum(xs, fs, nfft)

Spectrum of the first-differenced series, gain-compensated back to amplitude units. Differencing crushes what the polynomial detrend leaves below the gait band (slow multipath/body-shadow wander) — the dominant source of false cadence peaks — and dividing by the difference filter's gain |H| = 2 sin(pi f / fs) undoes its tilt so the peak location is unbiased. DC (gain 0) is pinned to zero, which the band never reaches.

F_gait_spectra(resid, fs, nfft)

(coincidence, full) magnitude spectra of the residual. The coincidence spectrum is the min() of the two half-window spectra: at ~26-sample windows and SNR ~1 a single periodogram's noise lobes can out-peak the true cadence line, but a noise lobe rarely lands on the same frequency in both halves, while the gait line is in both by construction. The caller picks the coarse peak on the min() and refines its location on the full-window spectrum, which has the sharper mainlobe.

Fincremental_cadence_hz(xs, fs_hz, band_lo, band_hi)

Firmware-shaped estimator: first-difference high-pass (kills the approach trend), Hann weight over the trailing window, Goertzel bank across the gait band. This is what a ring buffer plus a few dozen multiplies per block can afford on target; test_aliro_gait.py holds it to the FFT answer so the future firmware learner mirrors this exactly.

Fanalyze_walkup(label, idx, ranges)

Features from one approach's trusted-range series, already windowed to the descending phase by _approach_windows.

F_approach_windows(txn)

Split one session into its individual approaches.

When the phone keeps the BLE session up while the carrier walks up several times, one session.start..session.end holds multiple approaches, each ended by a relock.sent. Split on those relocks and keep each approach's descending phase (segment start .. closest approach), dropping the walk-away and the door-side loitering that follow it. Legacy single-approach captures (no relock, e.g. synthetic traces) fall back to the whole series, cut at the bolt stamp as before.

Fclassify(walkups)

Leave-one-out nearest-centroid over z-scored features. Returns None unless there are >= 2 labels with >= 2 analyzed walk-ups each.

F_scatter_svg(walkups)

Cadence vs stride regularity, one dot per walk-up, colored by label: the E1 eyeball plot (do the carriers cluster?).