Simulation of an autonomous vehicle with a vision-based navigation system in unstructured terrains using OctoMap

Abstract

Design and implementation of autonomous vehicles is a very complex task. One important step on building autonomous navigation systems is to apply it first on simulations. We present here a vision-based autonomous navigation approach in unstructured terrains for a car-like vehicle. We modelled the vehicle and the scenario in a realistic physics simulation with the same constraints of a real car and uneven terrain with vegetation. We use stereo vision to build a navigation cost map grid based on a probabilistic occupancy space represented by an OctoMap. The localization is based on GPS and compass integrated with wheel odometry. A global planning is performed and continuously updated with the information added to the cost map while the vehicle moves. In our simulations we could autonomously navigate the vehicle through obstructed spaces avoiding collisions and generating feasible trajectories. This system will be validated in the near future using our autonomous vehicle testing platform - CaRINA.