Jumping Jack Counter With Pose Estimation: Landmarks, Camera, Rep Logic

This page is the jumping jack layer under how rep counting works and how form feedback works: landmarks, camera placement, rep logic and what the two studies we read on October 18, 2026 report. Each number is one study's result under that study's conditions, so use it as a starting point to calibrate for your camera, model and users. This is engineering guidance for building a tracker, not exercise or medical advice.

Both studies come from clinical motor research and both used a Kinect, a 3D infrared camera that reports joints in three dimensions, rather than a 2D model running on a phone's RGB video. We use them only for how they set up the camera and defined the movement's signals.

The studies behind this page#

• Cho and colleagues (Scientific Reports, 2022) filmed 82 adults aged 18 to 65, 37 with autism and 45 controls, with a Kinect depth camera during nine movement tasks, one of which was 10 continuous jumping jacks.
• Hocking and colleagues (Journal of NeuroEngineering and Rehabilitation, 2022) tracked ten children and adolescents with autism, aged 10 to 17, with a Kinect V2 during six virtual-reality training sessions, recognised jumping jacks with a neural network, and measured each one.

Landmarks and angles#

In our reading, a jumping jack moves in the frontal plane: Hocking and colleagues define it as jumping from standing to a position with legs spread and arms raised, then back. Cho and colleagues measure it as two cycles, the hand cycle from maximal hand distance over minimal hand distance to the next maximal hand distance, and the matching feet cycle. Hocking and colleagues measure the highest wrist position as the maximum difference between wrist and shoulder heights divided by body height, and the widest leg split as the widest distance between the ankles divided by height.

Those signals need wrists, shoulders and ankles, which every keypoint scheme on this site has, from COCO 17 to BlazePose 33. The Kinect V2 skeleton Hocking and colleagues recorded has 25 joints and is not one of the maps on this site. In our reading, dividing by body height, as they did, is what lets one threshold work for people of different sizes and at different distances from the camera.

Where to put the camera#

Cho and colleagues mounted their sensor on a tripod at 1.4 m height with a vertical angle of −8°, with participants standing 2.5 m away. Hocking and colleagues placed theirs directly in front of the testing area, about 1 m away. Both set-ups face the person, which in our reading suits a frontal-plane movement: from the side, the hands and feet would move towards and away from the camera instead of across the image.

Cho and colleagues also controlled the scene: light, silhouette-fitting clothing, comfortable footwear, and an evenly lit environment free of distracting objects. One participant was excluded from their analysis for not meeting those filming conditions.

A rep state machine for jumping jacks#

In our reading, Cho and colleagues' cycle definition is a ready-made state machine: count one repetition each time the normalised hand distance passes a high threshold, drops below a low one and rises past the high one again, with the gap between the two as the hysteresis described in how rep counting works. Requiring the feet cycle to agree with the hand cycle would, in our reading, guard against arm swings without a jump. Cho and colleagues also report arrhythmicity, the absolute ratio between the standard deviation and the average of the cycle time, which a counter could use to flag irregular sets.

Hocking and colleagues recognised movements with a neural network over a sliding window of 60 frames, updating the result every 15 frames of new data, 0.5 s. Because they needed accurate start and finish times for each jumping jack, they checked every movement the network labelled as one and manually corrected the start or finish frames where they were wrong. They low-pass filtered the joint data at 12 Hz, shifted the coordinates so the pelvic joint sat at the origin, and cleaned each jumping jack's signals with a zero-lag Butterworth filter at a 3 Hz cut-off.

Failure modes the studies report#

• Boundaries. Hocking and colleagues' network found the movements, but its start and finish frames were not accurate enough for their analysis without manual correction.
• Noise. Hocking and colleagues filtered the joint data twice, the second time to remove high-frequency parts of the signals that the participants could not physically have produced and that came from the noisy extraction process.
• Filming conditions. Cho and colleagues' protocol depends on controlled light, clothing and background, and excluded a participant who did not meet it.

Pose estimation accuracy covers the general causes in more depth.

What was measured, and why it will not transfer#

In the text we read, neither study reports how accurately jumping jacks were counted or recognised, and no other study we read measured jumping jacks with a camera. Cho and colleagues cite an earlier technical validation of their protocol in healthy subjects, which we did not read. Hocking and colleagues describe their sample as very small.

The gap between these set-ups and a phone app is large: a depth sensor instead of an RGB camera, controlled clinical filming instead of a living room, and participant groups chosen for motor research. Build your own test set before you trust a jumping jack counter; testing a rep counting algorithm covers how.