Autonomous agents have been an arduous subject of research in Artificial Intelligence area. In order to develop some kind of autonomy in artificial agents, high level capacities like perception, reasoning and decision-making must be provided. The Situation Awareness (SA) field provides a theory that bases decision-making for agents, by performing the perception, comprehension and projection of their current situation. Thus, we aim to develop a SA model that generate believes about the autonomous agent’s situation in order to allow decision-making and action planning. The model will be applied in the scenario of autonomous vehicles, where the data provided by sensors may be affected by impreciseness or uncertainty in the measurements (e.g. the curve detected by sensors may be light or sharp). Besides, the phenomenon under observation is dynamic, since its varying with time. To address this issues, our model includes a hybrid inference based on Fuzzy-Bayesian inference and Dynamic Bayesian Networks.
A vehicle with a front camera performs a route on a highway. The images obtained by the camera are processed by a system through passive computer vision techniques to identify the path and possible obstacles. This information forms the basis of the agent’s beliefs, such as road and obstacles. If the road pattern changes considerably or the obstacle forms differ from its previous state, the agent will need to review its beliefs in those objects, that is, identify if what it believed to be, for example, a road is still a road. Thus, based on the agent’s beliefs, the stand-alone vehicle will make decisions such as staying on the lane or reacting to obstacles.
During the route, the features of the path are identified, such as the existence of curves, bifurcations, slopes and road interruptions. In addition, the possible obstacles, which include potholes, vehicles, pedestrians, animals and other objects are analyzed. If there is another vehicle, the system checks if it is moving and, in this case, what its direction is. It is not necessary that the agent has the ability to discern between pedestrians, animals and other objects that may be hindering the trafficability, because in these cases the expected behavior is the same.
Thiago Rateke is a Computer Vision Researcher with experience mainly focusing on visual perception for autonomous navigation. Finished his PhD degree at Federal University of Santa Catarina (UFSC) in 2020 with focuses on visual perception for Autonomous Navigation. Using approaches like: Stereo Vision, Optical Flow and Convolutional Neural Networks.