The 2-Centimeter Question: What Forensic Height Estimation from Surveillance Images Can and Cannot Prove

On the photogrammetric geometry behind a technique that performs better than most detectives expect, the biological variation that undermines every measurement if you ignore it, and the precise conditions under which height evidence drawn from CCTV footage is forensically useful versus forensically worthless.

There is a type of question that arrives in my forensic practice with reassuring regularity, usually from an investigating officer, occasionally from a defense attorney, and at least twice in my experience from a judge during testimony: the question of whether the person on the surveillance recording is 175 centimeters tall, or 180, or possibly 165, and whether I can say with certainty which of those applies, and whether that certainty would survive cross-examination. The question has a specific structure to it that tells me immediately how well the person asking it understands the methodology, because it conflates 2 distinct problems: the technical problem of measuring a pixel-image of a person against a reference, and the biological problem of what that measurement actually means as an attribute of a specific human being on a specific Tuesday morning at 03:17.

I have spent a significant portion of my professional life answering this question in one form or another, and the honest answer is yes, height estimation from surveillance images remains meaningful, frequently decisive, and occasionally misused in ways that could send the wrong person to court. Understanding which of those outcomes applies to a given case requires understanding both what the methodology does and what it cannot do.

What a Surveillance Camera Actually Records

The starting point for any forensic height estimation from images is an honest account of what a surveillance camera is and is not. A surveillance camera is not a calibrated measuring instrument. It is an optical system with a lens of specific focal length, pointed in a specific direction, at a specific angle of tilt, at a specific height above the floor, capturing light onto a sensor of specific resolution, with specific compression artifacts introduced by the recording codec, all of which are parameters that may or may not be documented, may or may not be retrievable from the installation records, and may or may not have changed since the installation was originally made.

When a person walks through the camera’s field of view, the image that results is a perspective projection of a 3-dimensional scene onto a 2-dimensional recording surface. Every meter of additional distance from the camera corresponds to a smaller apparent height of the subject in the image. Every degree of camera tilt introduces a systematic foreshortening. Every degree off-axis introduces keystone distortion. The image contains none of these parameters explicitly: it contains pixel values, and from those pixel values the analyst must reconstruct the geometry. This is the core technical challenge, and it is why “how tall is this person” is not a question you can answer by counting pixels.

How Photogrammetric Height Estimation Actually Works

The methodology that has become standard in forensic contexts, and that has withstood admissibility challenges in British, American, German, and other court systems, is photogrammetry, specifically the application of projective geometry to the problem of recovering metric information from images taken by uncalibrated cameras.

The foundational contribution in this space for uncalibrated camera applications is Criminisi, Reid, and Zisserman’s 1998 work on height measurement from video, which introduced vanishing point analysis as a means of recovering the geometric parameters of the scene from the image itself, without access to the camera’s technical specifications (Criminisi, A., Reid, I., & Zisserman, A., 1998, “A new approach to obtain height measurements from video”, Proceedings of the 1998 International Conference on Image Processing, 2, 879-883). The vanishing point of a set of parallel lines in the scene, typically the intersection of lines that are horizontal and parallel in the real world, allows the analyst to reconstruct the effective focal length and camera tilt that are implicit in the perspective projection. Once those parameters are known, a reference object of known height in the scene, typically a door, a counter, a vehicle, or an architectural element whose dimensions have been independently verified, provides the scale factor that converts pixel measurements to metric measurements.

The alternative approach is reverse projection photogrammetry, in which the analyst returns to the scene and places a human subject or a measuring instrument at the same position as the suspect, using a camera with the same geometry as the original surveillance installation, to reproduce the visual conditions and extract height by direct comparison. Both methods, when executed correctly with documented uncertainty propagation, are scientifically sound and court-admissible. The SWGDE Best Practices for the Forensic Use of Photogrammetry, updated in 2023, codifies the workflow for both (Scientific Working Group on Digital Evidence, 2023).

The research by Bieler and colleagues introduces the gravitational vertical as an additional geometric anchor that can be extracted from video without any reference objects in the scene, which is particularly useful when the camera angle is steep and architectural references are absent or undocumented (Bieler, D., Günel, S., Fua, P., & Rhodin, H., 2019, “Gravity as a Reference for Estimating a Person’s Height from Video”, arXiv:1909.02211). Ciampini and colleagues published in the Journal of Forensic Sciences in 2024 on combining video images with 3D laser scanning data to improve measurement accuracy and reproducibility in court contexts (Ciampini, V., et al., 2024, Journal of Forensic Sciences).

The Error Budget: Where the Centimeters Go

Any forensic measurement has an error budget, an accounting of all the sources of uncertainty that contribute to the difference between the measured value and the true value. For height estimation from surveillance images, this budget has several line items, and the honest practitioner states each of them explicitly in their report rather than presenting a single number as if it were precise.

The first line item is reference object uncertainty. If the analyst is using a door frame as the height reference, and the door frame is standard size, the uncertainty in that assumed height is typically on the order of ±1 cm, though it can be larger if the frame has been replaced, if there is a threshold step, or if the image geometry means the analyst is measuring to a point that is not the top of the frame but rather a lower feature that looks like it. Independently measuring the reference object after the fact is the only way to reduce this uncertainty.

The second line item is perspective and distortion. Even with a well-documented camera, the distortion introduced by the lens, particularly the barrel distortion characteristic of wide-angle surveillance lenses, redistributes pixel positions in a way that varies across the image plane. A person standing at the center of the frame will project differently than the same person at the edge. Software corrections exist and are effective when the camera’s distortion parameters are known through calibration, but many surveillance cameras have never been calibrated, and the coefficients must then be estimated or assumed.

The third line item is subject posture and foot position. The height that photogrammetry recovers is the distance from the top of the head to the floor at the subject’s foot position. A subject who is not standing erect, whose chin is lowered, whose back is curved, or whose weight-bearing foot is on a curb, a mat, or a slight grade will appear shorter than their true standing height. Footwear contributes: platform shoes, thick soles, and heeled footwear add height that may or may not be visible in the image. Gait phase matters: at the moment of heel strike in a stride, a person is measurably taller than at the moment of full foot contact.

The fourth line item is image resolution and compression. Camera resolution, specifically the number of pixels subtended by the subject’s height in the image, determines the minimum quantization uncertainty. A subject who occupies 50 pixels in height means that each pixel corresponds to approximately 3.5 cm on a 175 cm person, so the measurement cannot be more precise than that quantization limit regardless of the sophistication of the analysis. Compression artifacts, introduced by MPEG encoding at high compression ratios as is common in budget surveillance systems, blur the top of the head and add additional uncertainty to the exact pixel coordinates being measured.

Diurnal Variation: The 2 Centimeters Nobody Mentions in Court

Here is the part that nobody mentions in court, even though it is directly relevant to the comparison problem at the center of every height-based identification attempt, and even though the scientific literature on it is now several decades old.

Human stature is not a constant. It varies systematically across the day by a biologically determined amount that is larger than many forensic analysts report as their measurement uncertainty. The mechanism is well understood: the intervertebral discs of the lumbar and thoracic spine are viscoelastic structures that compress under the axial load of the body’s weight throughout the day, then recover during the unloaded rest of sleep. Eklund and Corlett documented a mean circadian variation of 19.3 mm in 8 young adults, with 54 percent of the diurnal loss occurring in the first hour after rising (Eklund, J.A.E., & Corlett, E.N., 1984, “Shrinkage as a measure of the effect of load on the spine”, Spine, 9, 189-194). Tillmann and Clayton found peak stature at 07:30 upon waking and the minimum at approximately midnight, with the maximum height loss by 15:00 of 1.44 cm (Tillmann, V., & Clayton, P.E., 2001, “Diurnal variation in height”, Annals of Human Biology, 28(2), 195-206). Studies in Ghanaian adults found an average diurnal loss of 1.61 cm with a maximum of 2.7 cm (Vuvor, F., & Harrison, O.A., 2017, Letters in Health and Biological Sciences, 2(2), 91-96).

The forensic implication is this: if a suspect’s police-measured height, taken at 14:00 on the day of their arrest, is compared with a height estimate derived from surveillance footage captured at 07:30, and if the analyst’s reported uncertainty range is ±2 cm, the diurnal variation alone can account for the full width of that range, meaning the biological variation is as large as the measurement uncertainty and the comparison is less discriminating than the reported figures suggest. This is not a reason to abandon height estimation. It is a reason to consider the time of day of the surveillance recording and the time of day of any reference measurement, to treat the resulting uncertainty as somewhat wider than the pure photogrammetric calculation suggests, and to document this consideration explicitly in the expert report.

What Accuracy Is Actually Achievable

A ±1.5% margin, which the SWGDE framework and forensic video specialists cite as achievable under good conditions, corresponds to ±2.6 cm for a person of 175 cm height. Under less favorable conditions, specifically lower image resolution, greater subject distance from the camera, uncertain reference object dimensions, or steep camera angle, this expands to ±5 cm or more. Combined with the ±1 to 2 cm diurnal variation, a realistic total uncertainty for a challenged height comparison in a typical surveillance context is often ±3 to 5 cm and occasionally wider.

This does not mean the measurement is useless. ±5 cm on a person of 175 cm is 2.9 percent, which is specific enough to be genuinely informative. It means that a suspect of 190 cm can be excluded from a frame that places the person at between 165 and 175 cm with high confidence. It means that a suspect of 172 cm cannot be excluded from the same frame, but cannot be definitively identified by height alone either. The measurement functions best and most reliably as a tool of exclusion.

AI, Deep Learning, and the Admissibility Question

The research trajectory in this field is moving toward automated, AI-assisted height estimation that does not require manual identification of reference points. Contemporary deep learning architectures, including pose estimation models that recover body landmark coordinates and monocular depth estimation networks, can produce height estimates from single-camera footage without requiring scene calibration. Finocchiaro, Khan, and Borji described egocentric height estimation approaches in 2016 using monocular cues (Finocchiaro, J., Khan, A.U., & Borji, A., 2016, “Egocentric Height Estimation”, arXiv:1610.02714).

The admissibility question for AI-driven methods is real and is not yet resolved in most jurisdictions. Traditional photogrammetry has a documented paper trail: the mathematics of projective geometry, the calibration protocols, the uncertainty propagation equations, the validation studies in peer-reviewed literature. An expert can explain in court, step by step and with algebraic clarity, how the reference point was identified, how the vanishing point was computed, how the reference distance was measured, and how the measurement uncertainty was derived. A deep learning network that produces a height estimate from a single frame is, from the court’s perspective, a black box whose output depends on a training dataset that may or may not be relevant to the specific subject, camera, and scene, and whose uncertainty is not expressed in terms that can be subjected to the same transparent scrutiny.

The SWGDE standard, the OSAC standards for forensic photogrammetry, and the European Network of Forensic Science Institutes guidelines all emphasize auditability and reproducibility as preconditions for court admissibility. AI methods that cannot satisfy these requirements in their current form will not be accepted in courts regardless of how accurate they appear to be in controlled tests, and accuracy in a controlled test with a known ground truth is not the same as reliability in a forensic context with an unknown subject and uncertain imaging conditions.

Exclusion, Inclusion, and the Distinction That Courts Must Understand

The most important conceptual distinction in forensic height estimation, and the one most frequently blurred in police reports and sometimes in expert testimony, is the distinction between exclusion and inclusion.

Exclusion means that a given height estimate, stated with its uncertainty range, is inconsistent with the suspect’s known stature, and that the person in the image is therefore not the suspect. This is the strong use of height evidence. If the surveillance image places the perpetrator between 160 and 170 cm with reasonable confidence, and the suspect stands 189 cm, the exclusion is robust and survives almost any reasonable challenge.

Inclusion means that the suspect’s known stature falls within the estimated height range and therefore cannot be excluded. This does not mean the person in the image is the suspect. It means that the suspect cannot be eliminated on the basis of height. This is a weak and frequently misunderstood form of evidence. Many people share the same height range. A height range of 170 to 180 cm encompasses a very large fraction of the adult male population of any European country, and stating that a suspect falls within this range contributes almost nothing to the identification probabilities in the case unless height is corroborated by other evidence.

The distinction matters enormously in practice because prosecutorial language frequently conflates “cannot be excluded” with “is consistent with the perpetrator,” and juries and occasionally judges interpret “consistent with” as far stronger evidence than it is. The expert witness has a responsibility to draw this distinction explicitly and to resist any formulation that implies more than the data supports.

The Software in Practice: What Analysts Actually Use

The theoretical framework described above is implemented in practice through a small number of specialized software tools that have been validated against known reference measurements and accepted by courts in multiple jurisdictions.

Amped FIVE and the related Amped Authenticate platform, widely used by police forces across Europe and North America, include photogrammetric measurement modules that implement the vanishing point approach with a graphical interface designed for analysts who are not optical engineers. The software guides the user through the process of identifying parallel line sets in the image, computing vanishing points, establishing reference objects, and propagating the uncertainty through the measurement chain. The output is a measurement with an explicitly stated confidence interval rather than a bare number.

PhotoModeler, developed originally for architectural and engineering applications, has been adapted extensively for forensic use and is the subject of several peer-reviewed studies examining its accuracy under controlled conditions. The key feature that makes PhotoModeler court-capable is not its precision in ideal conditions, which is excellent, but its complete documentation of every step of the measurement process, which allows a second analyst to reproduce the measurement from the same input data and obtain a consistent result. Reproducibility is a precondition for scientific credibility in court, and it is a precondition that far too many informal height estimates, conducted without documented protocols by investigators who are not trained photogrammetrists, fail to meet.

What the tools do not do, and this is an important limitation to understand, is relieve the expert of the responsibility to assess whether the scene conditions are appropriate for the methodology. An analyst who runs Amped FIVE on footage with a camera angle of 60 degrees from vertical, no visible reference objects, and a subject resolution of 30 pixels will get a number from the software. That number will be wrong, or more precisely, its uncertainty will be so large as to make it uninformative, and the software will not necessarily prevent the analyst from presenting it as if it were reliable. The responsibility for that assessment lies with the expert, not the software.

The German Courtroom Context: What StPO and Expert Witness Law Require

For German readers and for cases tried in German courts, a brief note on the procedural context is warranted. Under the German Code of Criminal Procedure, the Strafprozessordnung, expert witnesses are appointed by the court and carry an obligation of impartiality that is distinct from the adversarial expert-for-hire model common in the Anglo-American tradition. The German court-appointed expert is expected to present findings that are reproducible, based on recognized scientific methods, and accompanied by explicit statements of uncertainty. This procedural framework is, in principle, well-suited to the photogrammetric methodology described in this article, because photogrammetry does provide reproducible results based on peer-reviewed mathematics with explicit uncertainty propagation.

What German courts have occasionally struggled with is the question of what constitutes “expert qualification” in forensic video analysis, because there is no nationally standardized certification for this specialization in Germany comparable to the SWGDE certification in the United States or the CSFS certification in Canada. An expert appearing before a German court for height estimation from video footage should, at minimum, be able to demonstrate familiarity with the primary literature on the methodology, document their measurement protocol completely, provide an uncertainty statement derived from the actual measurement conditions rather than from a general “typical accuracy” claim, and be prepared to defend every step of the analysis under examination by both prosecution and defense counsel.

In practice, the quality of height estimation evidence submitted in German criminal proceedings varies enormously, from rigorous photogrammetric analyses with full uncertainty documentation to informal visual comparisons by investigators who have no specialized training, and courts do not always have the technical background to distinguish between the two. The expert witness in this field has a particular responsibility to be clear about what level of analysis was actually performed.

Does It Still Make Sense?

Yes, with conditions that the question itself implies but does not state.

Height estimation from surveillance images makes forensic sense when the image quality is sufficient to support the required precision, when reference objects of documented dimensions are visible in the image, when the subject posture is approximately erect and visible without significant occlusion, when the camera geometry can be recovered by vanishing point analysis or by return-to-scene measurement, when the resulting uncertainty range is calculated and reported honestly, and when the evidence is used for exclusion rather than as a stand-alone identification claim.

Height estimation from surveillance images does not make forensic sense when the image resolution is too low to resolve the subject’s head position within a margin narrower than the claimed uncertainty, when the camera angle is so steep that the subject’s true height is substantially foreshortened and reference object geometry cannot be reliably established, when the analyst treats the result as a point estimate rather than an interval, or when the evidence is presented as probative of identity without any other supporting evidence.

The addition of AI and deep learning methods to the analyst’s toolkit is a development to follow carefully, but not to deploy in adversarial proceedings until the auditability problem is resolved. Until validated against forensic standards and accepted by the relevant professional bodies, the methods articulated by Criminisi and refined by three decades of subsequent validation remain the appropriate standard for evidence that must survive cross-examination.

Height estimation from images makes sense exactly when it is executed with documented methodology, reported with honest uncertainty, and interpreted within the limits of what exclusion evidence can actually prove. That scope is narrower than investigators sometimes want, and more useful than a naive reading of the uncertainty ranges might suggest. Anyone who presents a 3-centimeter gap between the measured and actual value as evidence of methodological failure has understood neither the fundamentals of measurement science nor the biology of human spinal cartilage. Anyone who presents the same interval as proof of identity without further corroboration has understood neither the limits of the methodology nor the requirements of forensic evidence standards. Both errors are correctable, through training, through standardization, and through expert witnesses who maintain the distinction between what can be measured and what that measurement proves.

References

• Bieler, D., Günel, S., Fua, P., & Rhodin, H. (2019). Gravity as a reference for estimating a person’s height from video. arXiv preprint arXiv:1909.02211.
• Brolund, A., & Bergström, P. (2007). Estimation of human height from surveillance camera footage. Linköping University Electronic Press.
• Ciampini, V., et al. (2024). An innovative method for human height estimation combining video images and 3D laser scanning. Journal of Forensic Sciences. https://doi.org/10.1111/1556-4029.15378
• Criminisi, A., Reid, I., & Zisserman, A. (1998). A new approach to obtain height measurements from video. Proceedings of the 1998 International Conference on Image Processing, 2, 879-883.
• Eklund, J.A.E., & Corlett, E.N. (1984). Shrinkage as a measure of the effect of load on the spine. Spine, 9, 189-194.
• Finocchiaro, J., Khan, A.U., & Borji, A. (2016). Egocentric height estimation. arXiv preprint arXiv:1610.02714.
• Liscio, E., Guryna, H., Lea, Q., & Olver, A. (2021). A comparison of reverse projection and PhotoModeler for suspect height analysis. Journal of Forensic Sciences, 66(5), 1841-1855.
• Scientific Working Group on Digital Evidence (SWGDE). (2023). Best practices for the forensic use of photogrammetry. SWGDE Technical Document.
• Tillmann, V., & Clayton, P.E. (2001). Diurnal variation in height and the reliability of height measurements using stretched and unstretched techniques in the evaluation of short-term growth. Annals of Human Biology, 28(2), 195-206.
• Voss, L.D., & Bailey, B.J. (1997). Diurnal variation in stature: Is stretching the answer? Archives of Disease in Childhood, 77(4), 319-322.
• Vuvor, F., & Harrison, O.A. (2017). A study of the diurnal height changes among sample of adults aged thirty years and above in Ghana. Letters in Health and Biological Sciences, 2(2), 91-96.