Lucy Morecroft
Research student, Department of
Probability and Statistics
A Statistical Approach to
Forensic Facial Identification using 3D Images
Background
There
are several techniques already available for comparing and matching facial
photographs, however current practices used in this field are thought to be
unreliable, crude and unscientific. This undoubtedly leads to incorrect facial
matches or exclusions and mistakes like this can result in miscarriages of
justice. There are no recognised, tested scientific procedures for comparing
two facial images. There are a number of ‘experts’ in the area of facial
comparison, though there seems to be no general agreement in methodology or
whether procedures can even be justified scientifically in the actual world away
from controlled laboratory conditions. A reliable method of matching facial
images from technology such as CCTV would be invaluable to assist in criminal
analysis.
The
Data
The dataset for this project consists of a collection of 3D
facial images. Over 3000 different subjects volunteered to be photographed in
3D as part of the IDENT project, which was carried out by the
The 3D images were collected using a Geometrix®
Facevision 802 scanner. To see my face in the 3D
system click here. By dragging with the
mouse you can rotate the image in 3D.
Thirty points on the face (recognised
anthropometric facial landmarks1 – see fig.1) have been identified
on the 3000 3D facial images; this process was carried out twice for each image
to allow for the analysis of different sources of error associated with this
landmark data collection. The landmarks are facial features that correspond
between subjects, such as the corners of the eyes and the tip of the nose, so
analysis of these points can show how face shape varies between subjects.
Certain
factors mean that the data needs to be standardized to allow the direct
comparison of faces. Statistical techniques developed by Dryden & Mardia2
provide landmark-based shape
analysis routines where the data are aligned using Procrustes
analysis, which removes any arbitrary location, rotation and scale information to
preserve the shape. To investigate any structure
within the shape data multivariate analyses and classification techniques such
as cluster analysis can be used.
Facial Comparison/Matching
Recognized
techniques for presenting trace evidence to courts have been explored to attempt
to model face shape and calculate the probability of a facial ‘match’ between
two images. Aitken
& Lucy3 use Bayesian statistical methods and kernel density
estimates to calculate the probability that trace evidence (in the form of
glass fragments) came from the same source as the crime scene, and the
probability that trace evidence came from some other source in the known population
of trace evidence. These two probabilities are evaluated by means of a
likelihood ratio test. This approach has been applied to facial identification
in the sense that the test is a comparison of the probability that the face of
the suspect is the same as the face of the person committing the crime and the
probability that the face of the suspect lies somewhere else in the known
population.
References
1
Farkas,
L.G (1994) (ed.) ‘Anthropometry of the head
and face’,
2
Dryden,
3
Aitken,
C.G.G. and Lucy, D. (2003) ‘Evaluation of
trace evidence in the form of multivariate data’ Appl.
Statist, 53,
109–122.
Supervisor
Contact
Lucy
Morecroft
Department
of Probability & Statistics
Tel:
+44 (0)114 2223857
Email: l.morecroft
(at) sheffield.ac.uk