We present the latest results on learnable stochastic temporal models for automatic event and behaviour recognition in CCTV surveillance video. We introduce a novel approach to modelling and recognising complex activities involving simultaneous movement of multiple objects. Our approach differs from most previous work in that the visual understanding of activity is based on visual event detection and reasoning instead of object tracking and trajectory matching. Dynamic probabilistic graph models are exploited for modelling the temporal relationships among a set of different object temporal events. Typical applications of this technology include automatic semantic video content analysis, profiling and indexing of salient event and behaviour captured in CCTV video, and the early recognition of atypical behaviour in scenes where such behaviour could lead to a threat to safety.
Automated video monitoring of mobile objects is a growing trend in many sectors, especially in surveillance applications. Many research groups are addressing substantial efforts to develop autonomous applications able to recognize specified events. Reliability of such a system is mainly defined by its ability to extract and to track features of interest in image sequences. Since the capacity to perform this basic task is strongly related to changes in image
contrast, video monitoring units made up mono-spectral sensors offer limited performances in many situations. The best example of such a limitation is the uselessness of a visible CCD camera in low brightness scene. To overcome these shortcomings, we developed an acquisition unit including an uncooled VOx thermal camera (8-12 mm)and a High-Dynamic-Range-CMOS(R) color camera more suitable for outdoor applications. Unlike similar systems, we perform image registration at hardware level rather than at software level. Advantageous characteristics of such a design are presented in this paper. A simple framework is also proposed in order to achieve context-independent event extraction from color and thermal information.
The Police Scientific Development Branch (PSDB) is part of the Home Office. Amongst its other work, PSDB evaluates intelligent CCTV and vision-based detection systems (VBDS) for security and policing purposes. The principle aims of VITAL are to produce a collection of CCTV images that can be used to evaluate the performance of VBDS and also to provide typical real-world footage for research groups to use in their development programmes.
New chemistry for leuco fluorescin and leuco rhodamine for latent bloodstain and fingerprint detection has been developed in our laboratories. The use of these leuco dyes results in excellent contrast for several hours. The FBI's Evidence Response Team and DNA I unit collaborated with Georgia State University to validate the new fluorescin chemistry for use in the field.
In addition, several new NIR dyes have been developed in our laboratories that can be used to detect different chemical residues, e.g., pepper spray, latent fingerprint, latent blood, metal ions, or other trace evidence during crime scene investigations. Proof of principle experiments showed that NIR dyes reacting with such residues can be activated with appropriately filtered semiconductor lasers and LEDs to emit NIR fluorescence that can be observed using optimally filtered night vision intensifiers or pocket scopes, digital cameras, CCD and CMOS cameras, or other NIR detection systems. The main advantage of NIR detection is that the color of the background has very little influence on detection and that there are very few materials that would interfere by exhibiting NIR fluorescence. The use of pocket scopes permits sensitive and convenient detection. Once the residues are located, digital images of the fluorescence can be recorded and samples obtained for further analyses. NIR dyes do not interfere with subsequent follow-up or confirmation methods such as DNA or LC/MS analysis. Near-infrared absorbing dyes will be summarized along with detection mechanisms.
We describe an optical forensic method for tracing a CD back to the pressing machine in which it was created, and present a system we have developed which maintains a library of the 'fingerprints' of such machines and can compare sample CDs against this library. In principle, any security feature that is deliberately created can be copied by a counterfeiter. In our forensic method we concentrate on features that arise spontaneously in the manufacturing process. Such features act as a signature or 'fingerprint'. In the case of CDs we show how the moulding process leaves an imprint of an unpolished part of the 'mirror plate' on the CD surface. Using machine vision and pattern-matching, we demonstrate the use of the system to acquire a positive match of a sample against a pre-recorded library entry created using a different CD from the same mould.
An overview is presented of the history of retroreflection including biological and man made retroreflection and the evolution of the technology of the latter. Emphasis is on the optical characteristics of retroreflecting devices with the strongest emphasis on corner cube retroreflectors. A comparison of various corner cube and spherical retroreflector designs is given. Optical characteristics include return beam intensity, beam shape, beam polarization, effects from aperture sizes, materials and source properties such as coherent and incoherent light sources. The performance of retroreflectors having macroscopic vs. nanoscopic (wavelength scale) configurations and designed in error is discussed. Emphasis is on the performance properties of solid, open, full and partial corner cube retroreflectors made with various materials and coatings. An over view of a few applications for use of retroreflectors in today's counter terrorism and crime fighting world are discussed.
We describe the latest progress on two new terahertz (THz) prototype systems for security screening applications. Firstly, we show a terahertz stand-off detection system, working at a distance of 0.5 m, and with a dynamic range of ~60 dB. This allows the measurement of spectral features from 0.1 to 2.5 THz. Using this prototype, we present the reflection spectrum of an explosive material collected at 0.5m stand-off and compare it with the predicted spectrum. Secondly, a hand held wand using multiple terahertz detectors has been assembled. This resembles a metal detector wand in operation, and is designed for detecting explosives as well as both metallic and non-metallic weapons hidden under clothing or possibly in bags. The first data from this system is presented, showing that objects can be distinguished using their spectral features.
It's known that any iris recognition system is composed of four steps: iris image acquisition, iris texture isolation, features extraction, and classification. In the acquisition phase, our system begins with identifying the human iris automatically in a given image using neural networks to test whether it contains iris or not. If the iris exists in the acquired image it goes to the second step which is the iris texture isolation to localize the pupil boundaries, localize the iris boundaries, extract the iris texture, convert the iris texture to the polar coordinate system and then equalize the histogram of the rectangular iris texture. In the features extraction phase, our system uses the Haar wavelet to extract 72 features. In the classification phase, our system uses the matching ratio to identify or reject the subject.
We describe a camera capable of recording 3D images of objects. It does this by projecting thousands of spots onto an object and then measuring the range to each spot by determining the parallax from a single frame. A second frame can be captured to record a conventional image, which can then be projected onto the surface mesh to form a rendered skin.
The camera is able of locating the images of the spots to a precision of better than one tenth of a pixel, and from this it can determine range to an accuracy of less than 1 mm at 1 meter. The data can be recorded as a set of two images, and is reconstructed by forming a 'wire mesh' of range points and morphing the 2 D image over this structure. The camera can be used to record the images of faces and reconstruct the shape of the face, which allows viewing of the face from various angles. This allows images to be more critically inspected for the purpose of identifying individuals. Multiple images can be stitched together to create full panoramic images of head sized objects that can be viewed from any direction. The system is being tested with a graph matching system capable of fast and accurate shape comparisons for facial recognition. It can also be used with "models" of heads and faces to provide a means of obtaining biometric data.
We describe recent developments to a novel form of hybrid electronic/photonic correlator, which exploits component innovations in both electronics and photonics to provide fast, compact and rugged target recognition, applicable to a wide range of security applications. The system benefits from a low power, low volume, optical processing core which has the potential to realise man portable pattern recognition for a wide range of security based imagery and target databases.
In the seminal Vander Lugt correlator the input image is Fourier transformed optically and multiplied optically with the conjugate Fourier transform of a reference pattern; the required correlation function is completed by taking the inverse Fourier transform of the product optically. The correlator described here is similar in principle, but performs the initial Fourier transforms and multiplication electronically, with only the final most computationally demanding output Fourier transform being performed optically. In this scheme the Fourier transforms of both the input scene and reference pattern are reduced to a binary phase-only format, where the multiplication process simplifies to a simple Boolean logic XOR function. The output of this XOR gate is displayed on a state-of-the-art Fast Bit Plane Spatial Light Modulator (FBPSLM). A novel readout scheme has been developed which overcomes the previous system output bottleneck and for the first time allows correlation frame readout rates capable of matching the inherently fast nature of the SLM. Readout rates of up to ~1 MHz are now possible, exceeding current SLM capabilities and meeting potential medium term SLM developments promised by SLMs based on novel materials and architectures.
Atmospheric scattering causes significant degradation in the quality of video images, particularly when imaging over long distances. The principle problem is the reduction in contrast due to scattered light. It is known that when the scattering particles are not too large compared with the imaging wavelength (i.e. Mie scattering) then high spatial resolution information may be contained within a low-contrast image. Unfortunately this information is not easily perceived by a human observer, particularly when using a standard video monitor. A secondary problem is the difficulty of achieving a sharp focus since automatic focus techniques tend to fail in such conditions.
Recently several commercial colour video processing systems have become available. These systems use various techniques to improve image quality in low contrast conditions whilst retaining colour content. These systems produce improvements in subjective image quality in some situations, particularly in conditions of haze and light fog. There is also some evidence that video enhancement leads to improved ATR performance when used as a pre-processing stage. Psychological literature indicates that low contrast levels generally lead to a reduction in the performance of human observers in carrying out simple visual tasks. The aim of this paper is to present the results of an empirical study on object recognition in adverse viewing conditions. The chosen visual task was vehicle number plate recognition at long ranges (500 m and beyond). Two different commercial video enhancement systems are evaluated using the same protocol. The results show an increase in effective range with some differences between the different enhancement systems.
A data bank on prefrontal imaging under stressful conditions including deceit, has been gathered over several years on National and International populations using a contact imager pad consisting of 16 detectors and 4 sources, validating the concept of imaging prefrontal responses to stress, not only following the response of the PFC to imaging stress but especially of precognitive activations. We designed a new portable and non-invasive optical detecting system for remote sensing of deceit at 1~2m distance. The signals of pre- and post-cognitive function in deceit can be detected with very high sensitivity for blood volume and blood oxygenation detection at depths sufficient for PFC imaging and sensitivities of sub-micromolar oxy-hemoglobin and blood concentration detection. Thus, remote imaging of the process of decision making seems possible and examples will be presented using both contact and flying spot remote sensing.