Modeling human visual perception for target detection in military simulations

Abstract

The search and target acquisition models used in current military simulations for visual detection of ground soldiers are empirical. Although taking into account human performance data collected in field trials, they do not attempt to realistically model human search behavior. This, however, is necessary to achieve realistic target detection performance, including such phenomena as false positive detections at realistic locations. Working towards this goal, this research creates a model of human visual perception for the search of a human target. The contributions of bottom-up and top-down information on human visual perception are examined in a visual search experiment, which includes eye movement recording of the participants. The results show that semantically relevant scene information is used to guide the search process, influencing eye movements. Consequently, a predictive model of eye fixations is created which takes semantically relevant scene locations into account. These meaningful locations are extracted from ground truth simulation data and fused into a relevancy map. The relevance map is compared with eye fixations of participants searching for human targets in realistic scenes. This comparison shows that the relevance map predicts fixation locations very well. A combination of the relevance map with a salience map achieves even better prediction of eye fixations.