Review of Energy-Efficient Embedded System Acceleration of Convolution Neural Networks for Organic Weeding Robots

Abstract

The sustainable cultivation of organic vegetables and the associated problem of weed control has been a current research topic for some time. Despite this, the use of chemical and synthetic pesticides increases every year. This is to be solved with the help of an automated robot system. The current version of the weeding robot uses GPUs to execute the inference phase. This requires a lot of energy for an 8-track robot. To enable autonomous solar operation, the system must be made more energy efficient. This work aims to evaluate possible approaches and the current state of research on implementing convolution neural networks on low power embedded systems. In the course of the work, the technical feasibility for the implementation of CNNs in FPGAs was examined, in particular, following the example of a feasibility analysis. This paper shows that the acceleration of convolution neural networks using FPGAs is technically feasible for use as detection hardware in the weeding robot. With the help of the current state of research and the existing literature, the optimization possibilities of the hardware and software have been evaluated. The trials of different networks on different hardware accelerators with diverse approaches were investigated and compared.