Joint probabilistic modeling and inference of intersection structure

Many tasks relevant to driver support, semi-automated, and autonomous vehicles benefit from the use of high-precision street maps. Examples include vehicle detection and tracking, path planning, and control. Intersections present an unique challenge due to the lack of paint markings, increased number of valid (or semi-valid) vehicle behaviors, and increased variability in the trajectories actually driven - two vehicles driving the same left-hand turn may actually traverse significantly different paths. Here, we develop a probabilistic model and inference procedure for estimating the set of valid connections between lanes entering and exiting an intersection. The input to our algorithm is a set of vehicle tracks derived from Lidar data on an experimental vehicle. These tracks are both positionally noisy and often do not fully traverse the intersection due to loss of tracking for a variety of reasons. Furthermore, there is significant ambiguity in data association between the tracked vehicles and possible intersection structures. This motivates the use of a probabilistic model that can reason jointly over all the components of the intersection structure. Our algorithm is able to recover the intersection structure in a wide variety of situations and with noisy input data.