Learning driving patterns to support navigation decision making: Preliminary results

Navigation is a science and technology of finding the position, course, and distance traveled by a ship, plane or other types of vehicles. The navigation task is a decision-making procedure, where at least one of the input or output parameters has spatial properties, and is related to current or future vehicle position or attitude. In recent years research in car navigation systems and driver support systems achieved exciting results but there are still some outstanding problems. This paper presents a novel approach to providing driver support for guidance navigation tasks. Although driving is a very complex process, there are a number of regularities and signals from navigation sensors contain a large number of patterns that could be better exploited to support navigation guidance tasks. A system based on artificial intelligence techniques could learn common sequences of driving events (driving patterns) by comparing previous experience (pattern history) and current events (context). The system could predict future events and detect differences between predicted and actual signals. A detected difference could serve as the basis for providing support for guidance navigation tasks. In this paper we present preliminary research results. Driving and Navigation Navigation is a science and technology of finding the position, course, and distance traveled by a ship, plane or other type of vehicle. Stated simply, the problem of navigation can be summarized as answering the following three questions: Where am I? Where am I going? How should I get there? [Leonard, Durrant-Whyte, 1991]. As such this problem is common for all goal-directed movable entities, including humans (with and without vehicles), animals and autonomous robots. Using vehicles for transportation is a convenient way to improve speed or reduce strain and thus to enable much larger distances to be traveled. Advances in transportation provided the basis for advances in other aspects of human life, but also inevitably brought new problems to everyday human life. Increasing speed and traffic congestion become one of major death causes in the modern civilization. Most of the time during driving, drivers solve navigation tasks. A navigation task is a decision-making procedure, where at least one of the input or output parameters has spatial properties, and is related to current or future vehicle position or attitude. There are many navigation tasks, some of them executed independently from other tasks (like parking) and some of them synchronized with other navigation tasks. For the purpose of this paper, we could group navigation tasks into three layers of abstraction, as shown in Figure 1. Reactive navigation tasks (like stopping in the front of unexpected obstacles, turning away if we are too close to a kerb or an other vehicle) are simply reactions to perceptions without any planning and need little or no environment model at all. Medium-level navigation tasks (referred to as Guidance) require more details about the environment and our own position in it. They execute high level commands (drive along street, turn right, etc.). The other important goal for guidance task is to match the current position with the model of neighborhood and to send more accurate position information to higher level tasks. A planning task uses the global model data to plan future navigation steps in order to fulfill a previously set goal. Recently, advances in solid-state electronics, telecommunications and increased computing power lead to increased interest in using cheap navigation sensors [Kelly, 1995]. Development of global positioning system and availability of other inexpensive navigation sensors (like gyroscopes, accelerometers, range finders, etc) opened possibilities for more intensive research in car navigation systems. Most of the current research and commercial usage of navigation sensors are devoted to accomplishing higher-level navigation tasks by calculating vehicle position and attitude and comparing it with model of the environment and planned route. A typical example is shown in Figure 2, which depicts a global architecture for future car navigation and control systems as suggested by a number of leading European car makers and research institutions in the PROMETHEUS program [Catling, 1990]. Research in car navigation systems achieved some exciting results, and enabled development of many commercial systems. However, there are some outstanding problems: • The database of an environment model (roads, intersections, etc.) should be very large even if a driver does not need most of this data This database must be updated very often to be synchronized with changes in infrastructure. It is important to state that a need for a database poses significant problems (digital road databases are not available in majority of the countries, standardization, regular updates, etc.) • Data from sensors are fused into a very limited set of aggregate data describing vehicle position, attitude, speed, while a lot of other information from sensors are discarded. Figure 1. Navigation tasks and models.