Plank Detection With Pose Estimation: Landmarks, Camera, Hold Logic

This page is the plank layer under how rep counting works and how form feedback works: landmarks, camera placement, hold 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.

The studies behind this page#

• Sousa and colleagues (Journal of Functional Morphology and Kinesiology, 2026) counted 12 exercises, planks among them, with angle-threshold state machines on MediaPipe landmarks, and tested them on videos from public pose datasets, competition footage and their own recordings.
• Kim and colleagues (Frontiers in Public Health, 2026) built a browser-based MediaPipe Pose program with five PE teachers and five 9th-grade students over about three months at one public middle school, and report how its plank rule changed in classroom use. It is a qualitative case study, and the authors did not formally benchmark pose-estimation accuracy.

Landmarks and angles#

Sousa and colleagues model the plank as isometric exercise: a single holding state defined by torso angle thresholds, with an extra check that the body stays extended in a horizontal alignment throughout. Kim and colleagues' rules connect subsets of MediaPipe's 33 landmarks to conditions such as joint-angle ranges, alignment of body segments and temporal thresholds.

Neither paper names, in the text we read, the landmarks its plank rule uses. In our reading, a horizontal-alignment check needs points along the whole body line, such as the shoulders, hips and ankles, which every keypoint scheme on this site has. Foot points beyond the ankle depend on the scheme: BlazePose 33 and ML Kit have heels and foot-index points, while COCO 17 models stop at the ankle.

Where to put the camera#

No study we read tested camera placement for planks. Sousa and colleagues note that their pose model's performance can degrade when the subject is partially occluded or when camera angles deviate from frontal or lateral views. Kim and colleagues note that webcam-based 2D recognition can produce errors depending on camera angle, lighting conditions and body occlusion. The closest posture with placement data is the plank start of a push-up, and the push-up page covers which phone positions worked there.

A hold state machine for planks#

A plank has no repetitions to count, so the state machine tracks time. Kim and colleagues' teachers noted that activities such as planks emphasise sustained posture rather than repeated movements, and the team redesigned the plank rule from a simple hold condition into time-based tracking with continuous alignment checks. Maintaining a proper plank posture starts the timer, and an incorrect posture stops it and triggers feedback such as "Keep your back straight".

They also gate the start. Recognition that begins as soon as the program opens can misfire while learners are adjusting devices or moving into position, so their program shows guidance for a ready posture with recognition disabled, displays a progress indicator while the learner holds it, and starts tracking only once the posture is recognised with sufficient stability. In our reading that is a minimum-duration rule on entry, and the same idea on exit stops one noisy frame from pausing the timer. Sousa and colleagues use a related rule for dropped frames, keeping the last valid phase for typically 3 to 5 frames when no phase matches.

How long a posture must be wrong before the timer stops, and how far the torso may drift, are values neither study publishes for the plank, so set them from your own recordings. Kim and colleagues constrained their editable parameters to ranges derived from the pose-estimation model and the teachers' domain expertise.

Failure modes the studies report#

• Starting too early. Kim and colleagues saw misrecognition when recognition began while learners were adjusting devices or moving into position, which their ready-posture gate addresses.
• Off-axis views, occlusion and fast movement. Sousa and colleagues list these as conditions in which their pose model's performance can degrade, leading to incorrect or delayed state transitions.
• Camera angle, lighting and body occlusion. Kim and colleagues name these as sources of error for webcam-based 2D recognition, which they consider unlikely to match the precision of 3D motion capture.

Pose estimation accuracy covers the general causes in more depth.

What was measured, and why it will not transfer#

• Sousa and colleagues report 100% for planks, a minority class of n = 9 in their data, with a 95% confidence interval of 69.2 to 100%, and they describe results for such small classes as preliminary rule validation rather than definitive generalization.
• Their plank videos were 2 they recorded and 7 from public pose datasets.
• Kim and colleagues did not benchmark accuracy with precision, recall or related metrics, and did not retrain or benchmark the MediaPipe Pose model against ground-truth datasets.

Neither result tells you how a hold timer will behave with your camera, model and users, so record your own planks and check that the timer starts, stops and resumes when it should. Testing a rep counting algorithm covers building a test set.