Jadranka Sunde

An Experiment on Human Face Recognition Performance for Access Control

An experiment was conducted on human face recognition performance in an access control scenario. Ten judges compared fifty individuals to security ID style photos where 20% of the photos were of different people, assessed to look similar to the individual presenting the photo. Performance was better than that observed in the only other comparable live-to-photo experiment [1] with a false match rate of 9% [CI95%: 2%, 16%] in this study compared to 66% [CI95%: 50%, 82%] and a false reject rate of 5% [CI95%: 0%, 11%] compared to 14% [CI95%: 0.3%, 28%]. These differences were attributed to divergences in experimental methodology, especially with regards to the distractor tasks used. It is concluded that the figures provided in the current study are more appropriate estimates of performance in access control scenarios. Substantial individual variation in face matching abilities, response time and confidence ratings was observed.

On the application of a polling model with non-zero walk times and priority processing to a medical emergency-room environment

We consider a queueing model used previously (Scholz and Sunde, 1998) to model a military high frequency (HF) communication network. Our discussion has moved from the original case where we aimed to maintain a high grade of service for the highest priority traffic classes. The characteristics of this network were that link set-up time was longer than the service time: it often took longer to establish the connection than for the actual transmission of the message. In (Pearce et al., 2000) we looked at a polling model with multiple servers and multiple queues, with each server visiting the queues according to a server allocation algorithm. The queueing system comprises a set of waiting lines to which requests arrive to be served by a pool of servers. We restrict our attention to the case where movement of servers from queue to queue does not happen in zero time. Our discussion focuses on the basics of the two-queue situation. We look at the way this queueing model can be applied to a medical emergency room, where setting up for certain surgery procedures takes longer than the actual procedures themselves. The significance of this model is the applicability to different problems, from communication networks to medical emergency rooms.

Scenario testing methodology for the assessment of screening technologies

A range of advanced screening technologies has been developed to detect substances or objects that can pose a threat if carried on commercial aircraft. However, outside of technical testing in a laboratory setting, very little testing has been conducted to estimate their performance in an airport setting. This paper presents protocols developed by DSTO to assess the performance of security devices to detect contraband concealed on a person. These protocols are classified as scenario evaluations because they test the devices in a controlled environment that mimics crucial aspects of the airport environment (e.g., baggage and mock passengers) using trained operators. Such evaluations provide a range of data including reliability rates and timing estimates as well as detection and error rates which are very difficult to determine in either technical or full operational evaluations. The ability to detect and recognise dangerous objects at a safe distance is a very important task in a number of other defence, police and security applications and these protocols may be useful in other environments.

Infrared image enhancement and human detection performance measures

The ability to detect and recognise dangerous objects at a safe distance is a very important task in a number of defence, police and security applications. In this paper, we look at ways of increasing the effectiveness of infrared imagery for object recognition through processes such as super-resolution image reconstruction and deconvolution methods. In this paper, we propose two techniques for assessing image quality improvement: operator assessment and edge detection; and report on some initial work recently undertaken.

Additional obstacles in the implementation of computer technologies in small, non-English speaking countries

The purpose of this article is to point out some of the obstacles in the implementation of computer technologies that appear more often and are harder to solve in small, non-English speaking countries (Croatia as an example). Some of these problems are: language, character set, education, the problem of supporting science, staffing, the maintenance problem and the small market problem.

How to assess human visual performance on an operational task

A methodology is presented for the assessment of human operator performance in a detection and identification task, using two sets of infrared images of natural outdoor scenes with everyday objects used as targets. It includes measures of effectiveness such as operator detection rate, identification rate, false alarm rate, response time, confidence levels and image quality ratings. This robust methodology could be used in the evaluation of any image improvement technique or to evaluate different imaging techniques or technologies.

Improving infrared images for standoff object detection

The ability to detect dangerous objects (such as improvised explosive devices) from a distance is important in security and military environments. Standoff imaging can produce images that have been degraded by atmospheric turbulence, movement, blurring and other factors. The number and size of pixels in the imaging sensor can also contribute to image degradation through under-sampling of the image. Establishing processes that enhance degraded or under-sampled infrared images so that objects of interest can be recognised with more certainty is important. Super-resolution image reconstruction and deconvolution methods are explored, as well as performance improvement measures.

Kraft’s number and ideal word packing

In the noiseless context, it has long been known that the average encoded word length of an instantaneous or uniquely decipherable code can be made to lie between the source entropy and that value plus unity. We address the question of finding sufficient conditions on the code–word probabilities for it to be possible to make the average code–word length approximate the entropy by a smaller prescribed amount.