Why one sensor is not enough: measuring posture and movement second by second

Most wearable movement research uses one sensor. A wrist-, hip- or ankle-worn accelerometer is cheap, easy to put on, and good at what it was designed for: measuring the overall amount of activity, sedentary time or sleep. It is the basis of a large literature, and for those questions it works.

MAIJU uses four sensors, one on each limb, because it answers a different question. Not how much did the infant move, but what were they doing, second by second: lying on their back, on their front, sitting, crawling, standing, still or moving, and how skilfully. That kind of question has a different set of requirements, and a recent study measured them directly.

Posture is a whole-body configuration

An accelerometer measures two things about the place it is attached to: its orientation relative to gravity, and how it is accelerating. A gyroscope adds how fast it is rotating. Neither says anything about the rest of the body.

That is enough to count movement. It is not enough to tell postures apart, because a posture is a relationship between body parts. An infant sitting and an infant lying on their back can hold their arms in exactly the same way; a thigh looks much the same in crawl position as in prone. What separates them is how the arms and legs are placed relative to each other — which only a sensor on each end of the body can see.

Movement quality has the same problem. Telling rolling from pivoting, or wobbling in place from crawling, depends on how the limbs move together over a few seconds, not on the intensity at any one of them.

What the evidence shows

The question of how many sensors are needed, and where, was tested systematically on MAIJU recordings of 41 infants aged 4 to 18 months, with synchronised video annotated by trained human observers as the benchmark (JMIR mHealth uHealth 2025). The same classifier was retrained on reduced versions of the data — fewer sensors, lower sampling rates, accelerometer without gyroscope — and each version was scored against the video.

Agreement between the classifier and human video annotation, as Cohen's κ. Movement categories are harder for people too: two trained annotators agree on them at around κ = 0.60, against about 0.95 for posture (Communications Medicine 2022).

The authors' conclusion is plain: single-sensor configurations were not feasible for second-by-second posture and movement detection. The ranking was the same for posture and for movement — all four limbs best, then three, then one arm and one leg, then both legs, with arm-only and single-sensor setups last — and leg sensors did better than arm sensors. The minimum that still worked was one upper-limb and one lower-limb sensor.

It shows up in the summary numbers too

One might hope that second-level errors average out over a long recording. They partly do, but not enough. When the classifications were summed into the time spent in each posture, four sensors, and one arm with one leg, matched the human annotation closely for every posture (r = 0.96–0.999, except side lying at r = 0.73 for two sensors). A single leg sensor fell below r = 0.9 for four of the categories, and a single arm sensor below r = 0.8 for five.

The same held for the overall motor score. The BABA Infant Motor Score computed from four sensors was indistinguishable from the two-sensor version (r = 0.98), but single sensors reached only r = 0.78–0.79 against the reference, with individual errors of up to 45–50 points on the 0–100 scale. The errors were largest in exactly the period that matters most: the move from floor-based postures to sitting and standing, when an infant's repertoire changes fastest.

The gyroscope, and what does not matter

Two other design choices were tested. Dropping the gyroscope hardly affected posture, but hurt movement detection: with raw accelerometer data alone, four-sensor movement agreement fell from κ = 0.58 to 0.27. Rotation is much of what distinguishes one movement from another.

Sampling rate, by contrast, barely mattered. With four sensors, results were unchanged from 52 Hz down to 6 Hz, because the classifier reads the posture from orientation and the movement from its context over tens of seconds. That is useful in practice: a lower rate means smaller files, longer recordings in the sensor's own memory and a longer battery life.

Where actigraphy fits

None of this makes activity counts wrong. They measure intensity, and they are the right tool for questions about intensity. The large growth-chart study computed actigraphy-style counts from the MAIJU leg sensors alongside the posture analysis, and the two turned out to complement each other (Sci Transl Med 2026).

Split by posture, the counts told a clearer story than the total. Activity while in crawl position jumped around 9 months and activity while standing around 12 months — the ages at which fluent crawling and walking emerge — while activity while sitting barely changed across the whole age range. Leg activity during crawl posture tracked fluent crawling (r = 0.83), and activity while standing tracked fluent walking (r = 0.91). The study's own summary: activity counts become physiologically and developmentally interpretable once they are complemented by the context of postures and movement quality. And that context, the authors note, is only available from multiple sensors in fixed body positions.

Consumer fitness wearables seem to contradict this, since they recognise walking or cycling from a single wrist sensor. But they do it over minutes, and on long, rhythmic, repetitive movement. An infant's play is neither: video annotation shows postures and movements changing from one second to the next, and much of it is not rhythmic at all (JMIR mHealth uHealth 2025). A measurement that averages over minutes misses the behaviour it is trying to describe.

What this means for MAIJU

MAIJU puts four sensors on the upper arms and thighs, in pockets that hold them at a fixed position and orientation. Four rather than the minimum two is deliberate. Four sensors are the best-performing setup, and they add redundancy: a home recording where one sensor drops out still has three, and the analyses are trained to tolerate that. A recording is analysed when at least three sensors were working. Further, having symmetric placements open the possibility of studying movement asymmetry, which is an important clinical marker.

The same reasoning explains why our other garment, NAPPA, uses a single sensor. It answers a different question — sleep, breathing and body position through the night — where one sensor at the waist is the right tool. The number of sensors should follow from the question, not the other way round.

For the details of the classifier and its validation, see the posture and movement profile.