Step lab: a step counter you can watch think

What the detector does

This detector is ours, written for this page, and it is not a copy of any phone’s. It exists to make visible the choices every step counter makes somewhere. For each sample:

1. Magnitude. It takes the length of the acceleration vector, √(x² + y² + z²), so it does not matter which way the phone is held.
2. Low-pass filter. An exponential filter with a 3 Hz cutoff smooths away hand tremor and vibration faster than a walking rhythm. This is the green line.
3. Baseline. A much slower average (a 1.5 s time constant) tracks the resting level, gravity included, so the threshold is measured from wherever the signal sits.
4. Peak and threshold. A local maximum of the filtered signal is a candidate step if it rises at least the threshold above the baseline (default 1.2 m/s²). The dashed line is baseline plus threshold: a peak has to cross it.
5. Minimum interval. A candidate sooner than the minimum interval (default 300 ms) after the last counted step is ignored, so one footfall that rings twice counts once.

Every one of those numbers is a judgment. Raise the threshold and soft steps (a shuffle, a step on carpet, a phone in a loose bag) disappear; lower it and bumps that are not steps get counted. Lengthen the interval and a double impact merges into one step, until it passes your step period and every other step vanishes. The synthetic replay has 50 steps built in, three of them soft and one with a double impact, so you can see each failure on a trace whose true count is known.

Why your phone’s own count is different

The step counts your phone keeps do not come from a page like this. On Android, Google’s motion sensors guide says the step counter and step detector sensors are either hardware-based or software-based, and that the step counter has more latency, up to 10 seconds, but more accuracy than the step detector; both are covered, with Google’s wording, in TYPE_STEP_COUNTER vs TYPE_STEP_DETECTOR. On iOS, Apple describes what CMPedometer returns as system-generated live walking data: the step counts arrive already counted, as CMPedometer covers. Those pages quote what Apple and Google document about the APIs: what they return, when, and with what permission. They do not describe how the platforms turn motion into steps, and this page does not guess.

What follows from that is the practical point of the step counting API page: a step count is an algorithmic estimate from motion sensors, not a literal record of footfalls, so treat it as close rather than exact, prefer the platform’s aggregated total over summing sources yourself, and do not quote an accuracy percentage. Two apps that read different sources, a phone and a watch, or one counter that waits seconds to report and another that reports at once, will disagree for reasons this lab makes visible. A “today” total can also disagree for a reason that has nothing to do with motion: where the day starts, which is the bug in why “today’s steps” is a bug. Every phone motion API is indexed at phone motion sensors.

What the browser gives this page

• The devicemotion event, which MDN describes as fired at a regular interval; its interval property is the time between samples in milliseconds, which the page shows when the browser reports it.
• Acceleration in meters per second squared. The page reads accelerationIncludingGravity, which MDN says represents the raw data measured by an accelerometer and may be the only value available on devices that aren't able to remove gravity, and falls back to acceleration when that is missing. The baseline makes either work.
• On iOS Safari, a permission prompt. The page calls DeviceMotionEvent.requestPermission() inside the Start tap because, per MDN, it needs transient activation, meaning it must be triggered by a UI event such as a button click, and it resolves to "granted" or "denied". Browsers without the method need no prompt.
• Rounded values, on purpose. The W3C Device Orientation and Motion draft tells browsers to limit each acceleration value’s precision to no more than 0.1 m/s² and to expose the interfaces only in secure contexts, as fingerprinting defenses. This detector’s default threshold is 1.2 m/s², so the rounding does not hide a step from it.

On a desktop, or a browser that never delivers a sample, Start reports that nothing arrived within three seconds and the synthetic replay is the way to see the detector work.

Next step

Turning this into a choice of API:

• The Best Wearable Fitness & Health-Data APIs (2026) →
• Apple HealthKit Integration Guide (2026) →
• Health Connect API Integration Guide (2026) →