Constraint-Guided Automatic Side Object Placement for Steering Control Testing in Virtual Environment

Side objects are the common road objects placed alongside roadways, such as traffic signs, trees or street lights. Even though such objects do not obstruct a driving path directly, the visual perception-based autonomous features may recognize them in an unintended way impeding their ideal behavior. We propose a framework that systematically places various types of side objects in the virtual environment for testing the perception-based steering control. Firstly, we give the mathematical constraints that characterize both linear and non-linear geometric aspects in placing the side objects including the distances from a roadway or other side objects, and their placement patterns. Secondly, we define the placement distribution criteria that characterize how well the side objects are to be distributed along the roadways. Finally, our placement generation algorithm automatically determines the position of the side objects via the SMT (Satisfiability Modulo Theories) solver, and the generated placements are guaranteed to conform to both the constraints and distribution criteria. The experiment shows the scalability of the placement algorithm as to how fast a large number of side objects can be generated conforming to the aforementioned properties. In addition, we show how the Convolutional Neural Network (CNN)-based steering controller alters its behavior under multiple environments generated from our framework according to the RMSE and disengagement metrics.