Bicep Curl Counter With Pose Estimation: Landmarks, Camera, Rep Logic

This page is the curl layer under how rep counting works and how form feedback works: landmarks, camera placement, rep logic and what the studies we read on October 18, 2026 measured. 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#

• Chen and colleagues (Sensors, 2022) detected 12 exercises, biceps curls among them, from videos of 20 users, using an object detector to recognise the exercise and MediaPipe landmarks to measure a joint angle per exercise.
• Hwang and colleagues (Sensors, 2023) had 30 healthy adults with less than three months of exercise experience do five seated upper-body exercises, biceps curls among them, with dumbbells and resistance bands, and counted repetitions from OpenPose BODY_25 keypoints.
• Noteboom and colleagues (Frontiers in Sports and Active Living, 2022) filmed five young male athletes doing biceps curls and lateral flies with one camera, estimated 3D joint centres with a deep-learning model, and compared the muscle forces this produced with a marker-based system.
• Peres and colleagues (Journal of Functional Morphology and Kinesiology, 2025) filmed ten men doing barbell curls under three loads and tracked a marker on the bar rather than body landmarks.
• Fieraru and Goedemé (Journal of Imaging, 2026) benchmarked 19 monocular 3D reconstruction methods on fitness motions.

Landmarks and angles#

Chen and colleagues compute the angle at a node of interest from two other nodes, and for the biceps curl they place that node at the elbow; in our reading, the other two are the shoulder and the wrist. They then turn the angle into a completion value between the joint's initial angle when the exercise is ready and its final bending angle when it is completed. Hwang and colleagues skip the angle: they count from the periodicity of the left hand's position, using its vertical coordinate for curls, after dividing all coordinates by the length of the torso to remove the effect of distance and body size.

Every keypoint scheme on this site has shoulders, elbows and wrists. Points beyond the wrist differ: BlazePose 33 and ML Kit add pinky, index and thumb points, while COCO 17 and OpenPose BODY_25 end at the wrist. In our reading that matters for curls, because the hand holding the dumbbell is where the occlusion happens.

Where to put the camera#

The studies placed the camera in front or to the side. Hwang and colleagues put theirs 1.5 m in front of the seated participant at a height of 1.2 m, and Noteboom and colleagues 1.5 m in front at a height of 1.5 m, angled 15 degrees downwards. Peres and colleagues set the camera's optical axis perpendicular to the sagittal plane, a side view, and note that 2D analyses are prone to parallax and perspective errors.

Noteboom and colleagues point out that the curl moves mainly in the sagittal plane and that occlusion of the upper arm can be expected. In our reading, a front camera therefore sees the forearm swing towards the lens, which foreshortens the 2D elbow angle, while the hand's height still changes clearly; a side camera sees the elbow angle in its plane but, as Chen and colleagues found, detects only about half of the body's nodes.

A rep state machine for curls#

Chen and colleagues define a complete repetition by its completion signal: it rises from 0% to 100% and drops back to 0%, changing steadily and slowly. A signal that does not reach 100% or return to 0%, and changes unsteadily and fast, is an incomplete one, and a signal with almost no change is no movement. In our reading, that maps onto a two-threshold counter with a minimum time per phase, with the gap between the thresholds as the hysteresis described in how rep counting works.

Hwang and colleagues count the zero-crossings of the filtered hand height. Before counting, they pass the keypoints through a 15-sample moving median filter, which they use to cope with occlusions shorter than about 0.25 s, and remove the signal's average. Chen and colleagues apply a similar rule to the exercise label: it changes only after 15 frames, 0.5 s, of a different detection.

Failure modes the studies report#

• The equipment hides the wrist. Chen and colleagues found the wrist blocked by the fitness equipment, causing node errors, and saw only half of the body's nodes detected when filming from the side.
• Short or uneven sets. Hwang and colleagues counted correctly in long sets, where the periodicity was constant, but in sets with few repetitions the timing of each flexion varied and the count tended to be overestimated. Their method also produces high-frequency noise when the hand is occluded and is vulnerable to non-repeating, nonlinear movements with heavy weights.
• Someone else in the frame. Noteboom and colleagues lost three trials, 2%, because a second person was visible in the background, which produced impossible skeletons.

Pose estimation accuracy covers the general causes, and multi-person pose tracking covers keeping the right person.

What was measured, and why it will not transfer#

• Hwang and colleagues counted biceps curls and overhead triceps extensions with larger errors than chest presses, shoulder presses and seated rows; across the five exercises, the mean absolute counting error ranged from 0.88 to 3.21 repetitions.
• Chen and colleagues classified movements as complete, incomplete or no movement with 92.84% accuracy across all 12 exercises, and found complete and incomplete repetitions hard to tell apart because participants were not specifically asked to perform them.
• Noteboom and colleagues cite their earlier work, in which upper-limb joint angles from the camera method differed from a marker-based model by about 4 to 8 degrees RMSD on average.
• Fieraru and Goedemé found upright exercises such as standing dumbbell curls among the easiest for monocular 3D methods, at about 67 mm mean error, against 118 to 155 mm for the hardest ground-level poses.

None of these figures describes your app. Seated participants, front cameras and novices in one study, five young athletes in another: measure your own counter with your own users, as testing a rep counting algorithm describes.