The human brain is constantly overwhelmed with stimuli from the surrounding environment, and its processing capacity is limited. Nevertheless, it is able to perceive the surrounding environment in an efficient manner, by filtering out non-relevant stimuli and concentrating on the relevant ones. The attention mechanisms influence such filtering and are executed in a seamless fashion. Representing such attention processes in a computationally tractable way can be a daunting task. Indeed, no agreement seems to exist in neuropsychology and psychology literature as to how attention mechanisms are conducted, which may not have an exact computational counterpart. Current computational models of attention try to encapsulate those findings in several ways. However, no agreement seems to be found as well on the attention models. The dissertation proposes a framework in order to capture the visual attention mechanisms. Such framework is divided into two components. The first is focused on the definition of a testbed which can be used as a proxy of the real environment. The definition of a controlled environment, which ensures the safety of the participants and allows the extraction of several parameters from the behaviour of participants, is paramount. The second component is related to the construction of computational models of visual attention, which are responsible for assigning a saliency value to objects. Such value indicates the degree of which the object "pops-out" when a person is observing the environment. The models are composed by two components: the bottom-up component of attention model is focused solely on the properties of the visual stimulus; the top-down factor, modelled by a Gaussian mixture model, is based on data collected from participants immersed in a virtual environment. Participants were asked their opinion regarding aspects pertaining to the immersiveness of the virtual environment. Collected results seem to suggest that the virtual environment is realistic and responsive to actions performed by users. As for the attention model, the obtained results seem to indicate that a poor performance of the model is obtained, with 17% of the conducted observations with a correct classification of the most salient object. Moreover, the contribution of the top-down component is non-existent; in fact, the same performance is obtained when the topdown component is considered, thus acting as a redundant component. Such redundancy suggests that the mixture model acts as a surrogate of the attention model when looking for the most salient object in the scene. On the other hand, the obtained results suggest either the adopted modelling strategy is inadequate, or the chosen performance metric is insufficient, or the validation procedure is not adequate. Further developments are thus required. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement N. 723386 (SIMUSAFE).