Skip to main content
Translation updated
#human-height-estimation#monocular-vision#camera-calibration#object-detection

Estimating human height with a single camera

Three source papers illustrate the assumptions and error sources of geometric, regression and AI methods for estimating height with one CCTV camera.

A single CCTV image provides no absolute length reference. A person closer to the camera appears larger, so equal pixel heights can correspond to different physical heights. Monocular height estimation needs information about camera geometry, the ground plane, head and foot locations, and a reference for metric scale. An AI detector can locate the head and feet more reliably, but the geometric ambiguity remains.

Sources checked on August 25, 2026. The bibliographic details and figures below were checked against publisher or DOI records. Review the DOI for corrections and the dataset’s terms of use before implementation.

#Main findings

  • For a fixed camera with known or estimable height, tilt, and focal length, classical photogrammetry or a small regression model offers an interpretable estimate.
  • An AI detector can help under difficult surveillance conditions. Evaluate detection performance separately from height error.
  • The studies use different distances, field-of-view regions, poses, age distributions, and calibration methods. Their MAEs cannot be ranked directly.
  • Occluded feet, a person standing off the assumed ground plane, and uncorrected lens distortion can each introduce systematic bias.

#The projection equation

A pinhole camera projects a 3D point onto the image according to the following relation.

s [u, v, 1]ᵀ = K [R | t] [X, Y, Z, 1]ᵀ

K contains intrinsics such as focal length and principal point. R, t are the extrinsics relating world and camera coordinates. Head and foot pixels alone cannot determine height because both depth Z and physical scale are unknown. Each study below constrains these unknowns differently.

#Three studies compared

StudyInput and referenceMethodReported resultInterpretation
Lee et al. (2008)Fixed CCTV, reference points in the sceneProject a virtual 3D ruler after linear and nonlinear 3D–2D calibrationThe abstract reports greater stability during data convergence than earlier methodsA photogrammetry study. The earlier references to YOLO and OpenPose have been removed.
Li et al. (2015)Head and foot observations of people walkingFit focal length, tilt angle, and camera height by nonlinear regressionApproximately 1.39 cm MAE against ground truthAssumes a fixed, elevated camera with a small downward tilt.
Ratthi et al. (2024)Monocular surveillance video and a custom height datasetUse YOLOv7-OA and attention to detect occluded people, then estimate height by field-of-view regionThe paper highlights and conference summary report MAEs of 0.02–0.8 cm across regionsThe 1.39 cm figure comes from a different dataset. The two results do not establish a general ranking.

#1. Reference points and a virtual ruler in Lee et al. (2008)

The authors use reference points in the scene to calibrate the relationship between 3D space and the 2D image. They then project a virtual 3D ruler into the image. The measurement depends on on-site calibration and photogrammetry.

The method suits forensic analysis when investigators can obtain reference points in the scene. A moving camera or an inaccessible scene makes direct application difficult.

#2. Three-parameter regression in Li et al. (2015)

The method assumes a typical surveillance camera mounted high and angled slightly downward. It fits just three camera parameters to head and foot observations: focal length f, tilt θ, and camera height Hc. This avoids the sensitivity of vanishing-line detection to noise.

The result applies to the mounting conditions and observations in the paper. Camera shake, uneven ground, and occlusion at the foot–ground contact point can change the error.

#3. Improving detection in Ratthi et al. (2024)

The authors use a monocular camera and combine YOLOv7-OA with attention to detect people under occlusion. They examine camera height, deflection angle, field of view, age, and sex alongside the geometry.

The conference summary gives 0.02–0.8 cm MAE across field-of-view regions. Its two pages provide too little detail about the data split, sample size, repeated trials, and generalization to assess the result independently. These figures are reported by the authors.

#Sources of implementation error

  1. Detection error. How does the bounding box handle hair, shoes, or occlusion?
  2. Posture error. When a person bends their knees or walks, the vertical head-to-foot distance differs from standing height.
  3. Geometric error. Estimates of K, distortion, camera pose, or the ground plane may be wrong.
  4. Scale error. A reference length or camera mounting height may be inaccurate.
  5. Evaluation error. Assigning adjacent frames of the same person to both training and test sets can inflate apparent performance.

#Designing a reproducible evaluation

  • Separate training, validation, and test data by person and recording session.
  • Report MAE and bias by distance, field-of-view region, occlusion, pose, and age group.
  • Include a 95% interval and the within-person variance over repeated measurements alongside the mean error.
  • Run an oracle experiment with ground-truth head and foot locations separately from the end-to-end detector experiment.
  • Perturb the calibration parameters slightly and measure sensitivity.

#What the comparison leaves open

The available results do not identify a single best method for every setting. A fair next experiment would run the virtual ruler, three-parameter regression, and AI detector + geometry on the same video and data split. That comparison could show whether better detection actually reduces height error.

Connect