Individual model. Each synthetic individual gets a
parameter profile drawn once — body-size / amplitude scale,
stride-frequency offset, gait-regularity factor, and a collar-fit
posture offset. All of that individual's samples are generated from
their profile, so seeds vary the noise within an animal while
the profile varies traits between animals. Every row is tagged
with individual_id, so a classifier can be tested with a
leave-one-animal-out split, not just held-out noise.
Labelling & scoring. Every sample carries a
ground-truth label. Export Raw as the algorithm input and the
truth file (timestamp, sample_id, individual_id, label)
separately; hold the truth file back, run the algorithm, then join on
timestamp + sample_id to score.
Fidelity note. Procedural synthesis informed by the
biologging literature for each species (see references). Cheetah
parameters follow McGowan 2022 (behaviour ontology, features) and
Wilson 2013 (stride frequency, peak hunt accelerations); rotary
vs. transverse gallop mechanics follow Hildebrand 1959 and
Bertram & Gutmann 2009. Hyena parameters follow Minasandra
2023 for the behaviour ontology and ODR; gait amplitudes are
scaled from carnivore biomechanics
(Heglund/Taylor/McMahon 1974 for stride frequency; Biancardi
& Minetti 2012 for transverse-gallop differentiation) rather
than measured directly — no per-stride collar IMU study
has been published for Crocuta. Horse stride frequencies
follow Robilliard, Pfau & Wilson 2007; absolute neck-IMU
amplitudes are estimated from general equine biomechanics
(Robilliard et al. used limb-mounted sensors, not dorsal-neck
collars). Elephant parameters follow Hutchinson et al. 2003
(no trot/canter/gallop; the 'Groucho gait' at fast walk; no
aerial phase therefore no landing spike) and Soltis et al.
2012 (behaviour ontology including feeding, bathing and the
stereotypic sway). Feeding uses an explicit head-down base
posture so the gravity vector shifts realistically onto the
surge axis — a real, classifier-tractable signature.
Lion parameters follow Suraci et al. 2019 for the
behaviour ontology (rest / walk / stalk / charge) with
stride frequencies scaled DOWN from cheetah via
Heglund/Taylor/McMahon 1974 (~20% lower at 190 kg vs 50 kg)
and the ambush-predator activity pattern reflected in the
mixed-mode plan (long rests, short bursts). Dog parameters follow Chambers et al. 2021
(collar-mounted canine activity classification) and the canine
biomechanics literature; medium-dog stride frequencies are
scaled from cheetah via Heglund/Taylor/McMahon 1974, with
gallop kinematics drawn from the rotary-gallop family that
dogs share with the cheetah (Hudson 2011 compared greyhounds
to cheetahs directly). Human (hip-mounted) parameters follow the
activity-recognition literature anchored by Bouten et al. 1997;
step cadence at hip matches the canonical adult walking value
of ~2 Hz. Note that at hip, SIT and STAND are nearly
indistinguishable in pure static accelerometry — real
classifiers lean on transitions and micro-motion. Suitable for
algorithm development and regression testing; it does
not establish field accuracy. Calibrate against real
data where possible.