Towards Multi-Modal Risk Assessment

Two of the most known sources of danger for automated mobile platforms such as robots or automated vehicles arise from objects in the environment that they cannot perceive themselves directly and from the prediction of the environment, which is subject to uncertainties. We therefore introduce an approach to accumulate risks arising from various sources. The risk values of different sources are accumulated in two dimensional risk maps. To achieve risk-aware motion planning over the complete planning horizon, the risk maps are determined for each time-step resulting in a time series of risk maps. In order to determine the risk arising from occluded objects, two methods are presented to keep track of areas where occluded objects can be located in. The first method considers objects that follow a road network. The second method accounts for objects that move freely. As a third source of risks, the risk caused by the multi-modal predictions of the surrounding visible objects due to their uncertain intentions is included in the risk assessment. The approach is evaluated in simulated scenarios and on test drives with an automated vehicle in real world traffic. It was shown that the presented approach is able to accumulate the contextual knowledge about occluded areas to enable informed decision making, and that a more risk-aware behavior of the automated vehicle can be achieved.