Bird eye view map estimation for autonomous delivery vehicles with limited resources

Abstract

Autonomous vehicles require a thorough understanding of their surroundings to make decisions and move safely while driving. The task that provides this type of information is known as Scene Understanding and it has seen a growing interest in the literature in recent years. However, existing solutions are often prohibitive for industrial use due to challenging constraints such as heavy computational power requirements or expensive equipment. Furthermore, most of the research is carried out in the field of self-driving cars. Therefore, it is often difficult to use these solutions in niche fields such as autonomous deliveries. In this work, we aim to fill this gap. We propose a new deep learning algorithm focused on delivery robots that aim to generate occupancy grids of the surrounding environment directly from simple RGB sensors. The proposed model is designed to work in real time on systems with limited computational power and without the need for updates of existing models. We demonstrate that our proposed model surpasses a heuristic algorithm used by Starship Technologies and can generate accurate occupancy grids even in difficult situations such as at night or in rain.