Thanks for visiting Imaging and Machine Vision Europe.

You're trying to access an editorial feature that is only available to logged in, registered users of Imaging and Machine Vision Europe. Registering is completely free, so why not sign up with us?

By registering, as well as being able to browse all content on the site without further interruption, you'll also have the option to receive our magazine (multiple times a year) and our email newsletters.

Crime scene investigations

Share this on social media:

Topic tags: 

Greg Blackman looks at some of the imaging techniques used in forensics and criminal investigations, from the importance of CCTV to generating images of a suspect from an eyewitness account

The spread of CCTV in society has been widely documented, with the UK believed to have one of the highest numbers of cameras per head of population in the world. The introduction of CCTV surveillance in the UK in the 1980s was deemed a success, but the technology has come under criticism more recently for being ineffective – an internal Metropolitan Police report published in 2009 concluded that, in 2008, only one crime was solved in London for every 1,000 CCTV cameras. That said, police investigators still build cases on CCTV footage and images of the suspect captured on camera can provide valuable intelligence.

One of the problems with the current network of CCTV cameras that Dr James Orwell of the Digital Imaging Research Centre at Kingston University, London identifies is the lack of interoperability between systems. From a forensic or investigative perspective, matching a suspect caught on CCTV with images in a separate database or from separate CCTV footage takes a long time, because the systems are not linked.

‘The idea that there is one monolithic surveillance system monitoring society is completely unfounded,’ Orwell comments. ‘We have transport, local authority, police, traffic, highway systems, private sector cameras in shopping centres, airports and small shops, etc. Police tend to use all of these in a forensic investigation and there currently isn’t much of a system to allow them to interrogate all of these different resources of CCTV with any efficiency. It’s an extremely inefficient process.’

One of the projects undertaken at Kingston University was GENERICK (Generation, Encoding and Retrieval of CCTV-derived Knowledge), which aimed to address the lack of interoperability between CCTV systems. It formed the basis of Dr James Annesley’s PhD thesis and looked at aspects like the sort of metadata that needs to be stored alongside the original video and how this metadata would be represented. The project contributed to the MPEG-A part 10 standard, known as the Video Surveillance Application Format (VSAF).

Part of the problem is looking at how the systems talk to each other and exposing the appropriate interfaces so that they can be searched. Inexpensive analogue cameras often involve non-standard coding methods and Orwell says the capability to interrogate a system from other systems and link them is limited. ‘It’s a million miles away from something like the internet or the broadcast standards area, where it’s been in everybody’s interest for decades to try and make everything as interoperable as possible,’ he says.

Another standards initiative is ONVIF (Open Network Video Interface Forum), an open standard for the control of cameras and also for the way in which the data from those cameras is presented. Companies including Bosch, Panasonic, Sony, and Siemens are involved in the forum.

Eyewitness reports

Without any video footage of the crime, police have to rely on eyewitness accounts. Often the police will use facial composite software to generate images of a suspect from the witness’s testimony. One such software program is VisionMetric’s EFIT-V, which uses a genetic algorithm to create the facial appearance. VisionMetric is a spin-out company from the physics department at the University of Kent in the UK and its EFIT-V system is used by around 50 per cent of British police forces.

Giving an accurate description of someone you’ve just witnessed commit a crime is difficult to do. ‘The benefit of the EFIT-V software is that you’re moving away from the method of trying to describe individual features, because we know that people don’t find that easy,’ says Dr Christopher Solomon of the Forensic Imaging Group, School of Physical Sciences at the University of Kent and technical director of VisionMetric. ‘You can know someone quite intimately and you know what they look like but you can’t describe them adequately. Nor can you even necessarily identify the individual features in isolation – for instance, it’s difficult to pick out a nose from someone you know from 100 other noses.’

The EFIT-V system presents the eyewitness with nine computer-generated faces based on their initial description (25 year old, white male, thin, etc). The witness is then asked to respond to the faces by rejecting those that don’t look anything like the suspect and accepting those that bear some resemblance. Based on the witness’s decisions, the system produces another generation of faces with a preference to those features in the accepted faces. In this way, the software gradually learns through the responses from the witness what kind of face they are looking for and the computer-generated face gradually converges towards a face that resembles the suspect.

Two 14pt font ‘e’ characters captured with the DR CID hardware. Both show unique ink-based printing ‘parasitics’. Image courtesy of Hewlett-Packard Laboratories

The fundamental assumption of the EFIT-V program, according to Solomon, is that the model is sufficiently comprehensive – it’s got 12 different ethnicities and, obviously, two different genders – that any face can be made. However, the witness can’t be expected to alter each parameter to create the face. ‘When they change one of these parameters, they don’t necessarily have a perception of getting closer or further away from the face they want, because there are so many variables,’ Solomon says. The software essentially does this for the witness and builds up the suspect’s face based on facial composites that show some resemblance to the individual.

‘The witness can’t necessarily describe what they want but they will recognise it if they see it,’ Solomon continues. ‘It’s the same principle as saying “you look like so-and-so from the TV” – you can’t say what it is that reminds you of the celebrity, but there is a familiarity nonetheless.’

The program’s main application is for the police, although it is used for other things as well – it was recently used in a study on behalf of Disney to evolve beautiful and ugly faces to mark a new release of Beauty and the Beast.

VisionMetric also provides age-progression software that produces age-progressed images for, among other applications, tracing missing persons, specifically children. It uses statistical appearance models, originally developed by Professor Chris Taylor and Professor Tim Cootes at Manchester University, which are parameter-based models of a human face built around conducting principal component analysis (PCA), both with respect to the shape of the features and also the underlying texture or colouration. ‘These models essentially enable you to represent a face using a compact string of numbers – you can get a near-photographic representation of a face stored in about 50 numbers,’ explains Solomon. Feeding those 50 numbers back into the statistical appearance model will generate a face that looks near-photographic.

‘For age progression, the program is learning how those 50 parameters change when a face gets older. For example, there are certain common characteristics in relation to children undergoing puberty, for instance, which can be applied in all cases,’ states Solomon. By using a large sample of faces to build an appearance model and then from the age of each face the program is able to extract a kind of signature. Changing certain parameters by a certain amount in a systematic way will make the face look older.

Some of the soft texture effects of aging like wrinkling are carried out differently, basically by increasing the intensity of wrinkles as the face is presumed to get older.

Counterfeit goods

One of the tasks carried out by forensic laboratories is the identification of fraudulent documents and counterfeit goods. Hewlett-Packard Laboratories has developed an imaging system that could potentially be used by forensic investigators in the field. The high-resolution imaging device, known as the Dyson Relay CMOS Imaging Device (DR CID), can detect print marks and aberrations of significantly less than 10μm, the microscopic level at which a printed mark can be uniquely identified. The device includes custom forensic perimeter analysis software to determine the unique print characteristics.

The DR CID lens projects the image onto each 3.2μm pixel of the 3 Megapixel CMOS image sensor at a 1:1 magnification. ‘The lens is one refractive surface and a mirror,’ explains Guy Adams, HP senior research engineer at HP Labs in Bristol, UK, one of the lead scientists who developed the device. ‘It is therefore very cheap to manufacture – the lens assembly is not made up of numerous aspheric elements. This lens has the potential to move counterfeit document inspection out of the laboratory and into the hands of forensic officials in the field.’

The basis for developing the DR CID was inspection of print quality, but its uses extend to counterfeit inspection. The device can be used to inspect characters with less than a 1 in 109 chance of a false read.

The imaging device is a comparative solution; it detects microscopic abnormalities, but only compared to the original document. ‘Every printed character has a unique fingerprint,’ says Adams. ‘Even on a production line using the same printer and paper/cardstock, two letter “A”s will be different at a microscopic level due to random printing elements.’

The fibre structure of the substrate, firstly, is unique and the unpredictable interaction of the ink with the substrate makes each printed character distinctive. It is impossible with current printing technology according to Adams to reliably print a dot less than 21μm and thus random printing aberrations less than 10μm cannot be reproduced.

The DR CID can be used to check that a character is the one printed originally – it is a one to one reference. Adams states that the imaging device would not be able to directly identify the printer, although through matching the characters the machine used to print them could be traced. ‘The technique might be able to identify the class of printer, whether it’s an inkjet or a laser printer, for instance, but not reliably the make and model of the machine,’ he says. ‘However, the technique does provide much more information about the history of the document and the entire chain of custody for the item could be traced when the printed mark is matched.’ The device is still under development at HP Labs as a potential part of an authentication and inspection system. And while forensic laboratories carry out a range of services for the authorities, including document verification, the DR CID device holds the potential to identify counterfeit documents in the field.

Further information
Other tags: 

Related features and analysis & opinion

22 February 2019

Ron Low, Framos head of sales Americas and APAC, reports from Framos Tech Days at Photonics West in San Francisco where Sony Japan representatives presented image sensor roadmap updates