Cooperative crowd mapping for interconnected autonomous vehicles

点击翻译按钮获取中文摘要

The goal of this project is to develop algorithms which derive consistent models of the traffic area and typical movement patterns of traffic participants from data that were collected by many vehicles over an extended time period. The derived information will be provided to other cooperative vehicles so that they gain better knowledge of the road topology and potential hazard areas. Furthermore, the algorithms will be able to detect systematic changes over time and to update their knowledge incrementally.Two types of sensor input will be used in this project, (a) the trajectories of cooperative vehicles (calculated from GNSS measurements and stereo camera image sequences), and (b) trajectories of other traffic participants (pedestrians, bicyclists, tramways, cars) which are observed by the cooperative vehicles using stereo cameras. The project does not use expensive sensors (e.g. highly accurate multi-layer lidar sensors) since it is not likely that these will be part of cooperative vehicles in future.The following information will be extracted from the observed trajectories: the number and layout of lanes, the topology of, and possible maneuvers at intersections, stop lines, zebra crossings, and points where pedestrian are likely to cross. Beside these static aspects, the approaches will be able to detect deviations from previous knowledge like blocked lanes or an increased density of pedestrians.Both, descriptive and predictive models will be used to describe the traffic area and movement patterns. As descriptive models we will use semantic maps in which the traffic space is represented using a formal grammar. For example, the semantic maps will contain the layout and the typical movements patterns at intersections. The predictive models will be used to predict movements of traffic participants. For example, they will allow to predict the future trajectory of an observed pedestrian. Like in the case of the semantic map, each element in the predictive model will be attributed with the position in the traffic area to which it refers, i.e., each intersection will have its own pedestrian behavioral model.For a comprehensible evaluation of our approaches, we will create benchmark datasets for cooperative crowd mapping. These will be provided online to other research groups.