long wait for a non-comedic television series about Cambridge
University is finally over. The scientists are sexy, the data is messy,
and without a doubt, the evidence never lies. In fairness, we don't
have the same camera crews and aren't even consultants to the show's
producers. We should however, hold our heads high, knowing that our
research is advancing the state of forensic science. The glitzy shows
are about us, though we may not always realise the extent.
Although science-fiction inspires many inventions, my interest in crime scene forensics started two years ago at a linguistics talk about speaker identification, an application I hardly cared for. Dr Gea de Jong, now a Phonetics researcher in our Faculty of Modern & Medieval Languages, spoke on her use of pattern analysis (my reason for attending) in speech forensics. She told anecdotes of using a recorded message to confirm a suspect's involvement in a hoax bomb threat at the Royal Albert Hall. It turns out that an expert can sometimes identify a person's regional birth-accent and the approximate number of years that individual has been living in a new city.
She gave other fascinating audio-analysis examples, such as one involving recordings of gun shots. A 911 recording of a Florida gas-station robbery helped verify which of two guns was fired by the owner, providing crucial evidence of self-defence. The guns were different models, but were not obviously identifiable because of telephone distortion and different locations relative to the handset (the louder shotgun was fired from the doorway).
There are too many crime-scenes, not enough experts, and most critically, evidence must be examined objectively in a repeatable fashion to be useful - and convincing. Fingerprints, first described by Prof Purkyne at the University of Breslau in 1823, are the oldest biometrics used to consistently identify individuals. Among the newest is iris recognition, developed by Prof John Daugman at our Computer Laboratory. His patented algorithm is in general use for identity verification, and was notably used to confirm the identity of the National Geographic cover's Afghan girl 18 years later. Both the cops and robbers now realize that a fingerprint or a photograph of an eye leads directly to a suspect's identity if they are already listed in an accessible database.
These techniques are proven, and some good databases exist, but I don't work on fingerprints, irises, or audio signals. The pattern analysis in Dr de Jong's talk demonstrated that I and many other scientists are already doing research that could be applied to crime scenes. So much of science is focused on modelling patterns, and investigators on the front lines cannot know that what I am studying could help them. In turn, crime scenes can motivate new research problems for us. Blood spatter analysis is one example.
Some scenes of violent crime contain blood stains. Blood spatter stains occur when blood flies off passively due to force being applied to a body. In 1895, Prof Piotrowski at the Jagellonian University was the first to propose that the elongated shape and layout of the stains indicated the location of a victim's head at the time it was subjected to trauma. There is a well established technique by which a specially trained forensic technician measures the individual blood spots (usually 100s of them) with a ruler.
The stains are affected by many physical variables, such as speed, liquid density, and the material properties of the surface. Air resistance affects the otherwise spherical shape of an airborne blood drop only slightly. Ideally, once the sphere lands on a flat surface, the collision flattens the liquid into an ellipse. The proportion of the ellipse's width to length reflects the angle of impact.
Currently, the technician pins a string to each spot, stretching it across the room to approximate the projectile motion with a line. This eventually gives the forensic expert a good idea of where the victim stood on the floor plan when hit (the strings intersect in the same area), and a rough idea of the height where impact occurred. Working with Amy Shen, a recent CU Engineering graduate, our main contribution is an algorithm that processes digital images of the crime scene to obtain the same information as the current "string method." Our experiments (so far, only using red paint) indicate that the algorithm matches the accuracy expected from a forensic investigator. Our secondary contribution is the exploitation of calibration objects to perform image rectification, producing shot-from-above images of the whole crime scene.
Forensic crime investigators will benefit from further developments in pattern recognition, AI, and the modelling of natural phenomena. Shoe prints are still photographed and posted on web-sites in hopes that someone will recognize the brand and size. Forensic artists still sketch suspects instead of adjusting photo-real models. Age progression is a black art. Polygraph tests are STILL used as supportive evidence. There is no Bat-computer that determines a sample's chemistry.
Specialists in other fields and investigators working on real cases don't always know what tools to ask for. As scientists, we are in a great position to examine the utility of our findings, but we have to talk to those "outsiders." Responsible application of science can coexist with the catching of bad guys and the exoneration of good guys.
Note: Gabriel's work was done in the Computer Vision Group of Prof Roberto Cipolla. Very few undergraduates were harmed in our experiments.