Scene Context-aware Rapidly-exploring Random Trees for Global Path Planning

This paper introduces a global path planning method for autonomous systems. Global path planning finds a feasible and collision-free path in an environment in which various kinds of regions and objects exist. However, the most planning methods use information such as collision-free space and obstacles in the environment. Interactions at each region (e.g., sidewalk and pavement) would be different. In this paper, we propose a method for global path planning taking semantic scene context into account. In contrast to conventional path planning methods which use collision-free and obstacle regions, the proposed method represents an environment as a cost map. The cost map is estimated from demonstrated human behaviors and feature maps derived from semantic scene context. To find a path on the cost map, we define a path cost and leverage an optimal rapidly-exploring random tree (RRT*) algorithm. We evaluate the proposed method regarding accuracy and computational efficiency with two public datasets and our contributed dataset. Experimental results show that our method successfully reproduces paths like human behaviors in short computational time.