Networked Edge Intelligence for Autonomous Farm Vehicles

Abstract

Precision agriculture and aging farming populations have given rise of autonomous farm vehicles to help reduce labor cost, save agriculture resources, and improve farming productivity. Unlike self driving on city roads or in dedicated facilities, autonomous vehicles on farmlands face a very different set of challenges. For example, large farms lack network coverage, the dynamics of the farm landscape change, soil conditions, and farm vehicles can be affected by weather and working conditions. Thus, autonomous farm vehicles require strong on-board sensing, cognition, and adaptive real-time decision-making capabilities. In this article, we analyze the autonomous capabilities for farm vehicles by determining general requirements, then pointed out the specific capabilities. In addition, we explore a multi-dimensional network with space-air-ground integrated network (SAGIN) architecture for autonomous farm vehicles that incorporates key technologies to address the needs of comprehensive network coverage, flexible networking, and large-scale access. Besides, the primary challenges faced by edge computing are addressed and we conduct experiments on diverse multi-models on various GPU platforms. The results show that our edge intelligent based AI inference scheduling framework as a significant overall performance gain on different GPU platforms compared with the baseline inference frameworks. Furthermore, we analyze the challenges of autonomous farm vehicles.