The Performance of a Biologically Plausible Model of Visual Attention to Localize Objects in a Virtual Reality

摘要

Visual attention, as a smart mechanism to reduce the computational complexity of scene understanding, is the basis of several computational models of object detection, recognition and localization. In this paper, for the first time, the robustness of a biologically-constrained model of visual attention (with the capability of object recognition and localization) against large object variations of a visual search task in virtual reality is demonstrated. The model is based on rate coded neural networks and uses both bottom-up and top-down approaches to recognize and localize learned objects concurrently. Furthermore, the virtual reality is very similar to real-world scenes in which a human-like neuro-cognitive agent can recognize and localize 15 different objects regardless of scaling, point of view and orientation. The simulation results show the neuro-cognitive agent performs the visual search task correctly in approximately 85.4 % of scenarios.