Risk assessment for autonomous vehicles in occluded areas: Identifying and mapping potential risks originating from occluded areas

Abstract

Similar to an experienced human driver, an automated vehicle needs to take the unknown into account while driving. This work proposes a way to perform risk assessment for autonomous vehicles when approaching areas that are occluded from the vehicle’s sensor. The approach uses an occupancy grid and map data to describe risks associated with occluded areas. A risk assessment algorithm evaluates how occluded road users may pose a risk to safe driving, using the reachability of occluded road users through the use of phantom objects (e.g. objects that may be in an occluded area, with associated speeds, directions, etc.). The phantom object’s predicted movement is modelled with an implementation of the A* algorithm, and this work incorporates a cost point in the A*’s heuristic function, to create realistic turning trajectories. The reachability is expressed in a grid as a probability of reach given a time horizon. The method may help automated vehicles evaluate the potential existence and movement of phantom objects to heighten their perception when navigating through an environment with occluded areas. Evaluation of the method was made on data collected from real traffic scenarios and self-created synthetic data depicting occluded scenarios in urban environments. Qualitative analysis of the results show how the method reliably finds occluded phantom objects and predicts reasonable trajectories of their movement. Successfully assigning the probability of a reach for each cell in the grid that a phantom object’s predicted trajectory traverses. Key