Header for SPIE use Detecting Hidden Targets: A Procedure for Studying Performance in a Mine-Detection-like Task Daniel T. Cerutti**, Ioan M. Chelaru, and John E. R. Staddon DukeUniversity, Psychology: Experimental, Durham, NC, 27708 ABSTRACT We report preliminary results from an experiment designed to study the perceptual and learning processes involved in the detection of land mines. Subjects attempted to identify the location of spatially distributed targets identified by a sweeping a cursor across a computer screen. Each point on the screen was associated with a certain tone intensity; targets were louder than “distractor” objects. We looked at the effects on target detection and false-alarm rates of the intensity difference between target and distractor signals, the number of distractors and training order. The time to detect 50% of targets (threshold detection time) was measured by a rapid adaptive technique (PEST) which generated reliable thresholds within few trials. The results are consistent with a simple model for the detection of cryptic prey by foraging predators: search was slower with more distractors, and the effect of distractors was greater when intensity ratio (IR) was lower. Although subjects got no accuracy feedback, performance improved somewhat with experience and was better in the low-IR condition when it followed the high-IR condition. The procedure seems to be a useful one for studying more complex mine-related detection tasks with a range of signal types and numbers of concurrent detection signals. Keywords: land mine detection, simulation, threshold, adaptive staircase, training, foraging, cryptic 1. INTRODUCTION Land mines are effective barriers to the extent that they can be concealed. Development of sensitive metal-detection equipment reduces concealment, but has led to the evolution of low-metal or no-metal mines, which pose new technical problems. Mine detection will always be subject to technical limitations. Mines must usually be detected by human operators using hand-held devices. Improving the performance of the human operator in the man-machine system is therefore just as important as perfecting the technology of detection. Testing system performance in a naturalistic setting – real mine lanes, real mines – is expensive and difficult. The human operator soon learns the characteristics of the setup, at which point data become less useful. Changing the setup is expensive and time consuming. Two ways to deal with this problem are (a) to develop a “virtual minefield,” a computer-based system that closely matches the natural one in terms of the tactile, visual and auditory information presented to the operator. Such a system would allow the effect of technical and training modifications on operator efficiency to be tested directly. And (b) to learn more about the basic psychological processes that underlie operator performance in mine-detection tasks. Option (b) is less direct and less certain to succeed than option (a), but may suggest specific technical changes and may have long-run benefits in adding to basic science. We now report preliminary results from a project aimed at understanding the psychological processes involved in mine-detection-type tasks. An Analogy from Behavioral Ecology Most laboratory studies of detection involve trial-by-trial procedures and allow the test subject little freedom of action. Their focus is on the sensory limitations of the subject not on the way the subject interacts with his environment. But mine detection is more like the problem faced by a predator attempting to find a concealed or cryptic (camouflaged) prey than a sensory-discrimination task. There is a substantial amount of research on the detection of cryptic prey4,5,7,8,13. Relevant work has looked at how rate of target (prey) detection depends on the crypticity of the target. For example, Gendron2 studied in a semi-natural environment how quail forage for food objects that matched (high crypticity) or did not match (low crypticity) their background. He found (a) that the animals rarely made a mistake (false alarm rate was close to zero); (b) they search faster for the less cryptic food. Based on these results, Gendron and Staddon3 proposed a descriptive model for how detection of cryptic prey depends on crypticity and search rate. Figure 1 shows the suggested relation. Probability of detection remains high with conspicuous (low crypticity) targets at all but the highest * Correspondence: Email: [email protected]; Telephone 919 660 5707; Fax 919 660 5726search rates. With less conspicuous (cryptic) targets, probability of detection decreases rapidly as search rate increases. These relationships can be summarized in the following equation: PKd + (S/M)K = 1 or, Pd = [1 - (S/M)K]1/K, K > 0. (1) where S is search rate, M is the maximum search rate, K is a measure of crypticity, or how well the target is masked by extraneous noise, and Pd is probability of detection. The model predicts that the probability of detecting a target, Pd is a joint function of both target crypticity and search rate, S. Moreover, the optimal search rate (at which rate of target detection is maximized) decreases as targets become more cryptic. Figure 1 illustrates a further assumption of the model, that all curves intersect the abscissa at the maximum search rate (M). This represents an upper limit on the rate of search imposed by the observer’s perceptual or physiological abilities or by the detection equipment. In more complex detection tasks, with two target types that differ in crypticity, the optimal search rate will also reflect their relative densities3. Gendron and Staddon4 did an experiment to test this model with human subjects. Subjects searched for computer-generated visual targets. The targets were characters embedded in a background of nontarget symbols (“distractors” or “clutter”). Crypticity was manipulated by varying the number of distractor characters and search rate was manipulated by varying the display duration. The main results were (a) that probability of detection, Pd, decreased as search rate increased; (b) the decrease in detection probability was more rapid for more cryptic targets,; and (c) that Pd tended to converge at one point on the abscissa, the maximum search rate1. All these results are in agreement with model predictions, which encouraged us to