A theory and model to explain the variance in performance of visual search for ground vehicles in natural terrain
This research developed a theory and mathematical model explaining the variation in observers' visual search performance that can be attributed to computable properties of the image stimuli. The specific task is search for stationary military vehicles on natural terrain in narrow field of view images taken from the perspective of an observer on the ground, with vehicles in tactically appropriate deployment, including the use of camouflage, cover and concealment. Search performance refers to the joint distribution over the observer population of search time and outcome: detection, or quitting without detection. The research involved secondary analysis of data from two large scale experiments, a search experiment and a cued detection experiment, each using the same set of 1150 images. The search theory explains the dynamic interaction of scene learning and target seeking activities, the dynamics of termination, and the origin of systematic differences in the patterns of performance among observers. The theory predicts that the distribution of search time has a characteristic shape, and that search is a race between quitting and detection. A mathematical race model using the convolution of two unequal negative exponential distributions explains the empirical search performance data. Other quantitative implications of the theory are also substantiated by the data. The theory and computational model address the visual attributes that make a substantial contribution to observers' ability to detect a vehicle. The novel and significant features include methods to represent the adaptive organization of target perception, the size- and shape-adaptive local surround reference for region segmentation in inhomogeneous surroundings, and the functional relationship of size, contrast, clutter, shape evidence and color of the component target regions to detection performance. The model accounts for 72 percent of the variance in detection performance over a large and diverse set of target and image conditions. When the data are pooled over images of essentially the same scene with minor differences to reduce sampling error, the proportion of variance explained by the model increases to 77 percent. The model is unbiased with respect to scene luminance, color versus monochrome images, scene haze, and vehicle scale or range.
|Year of publication:||
Wayne State University
|Type of publication:||Other|
ETD Collection for Wayne State University
Saved in favorites
Similar items by subject
Find similar items by using search terms and synonyms from our Thesaurus for Economics (STW).