First Results of a Comparison of Machine Learning Hardware Acceleration Approaches Using Field Programmable Gate Arrays in an Agricultural Mobile Robotic Application Case

Demanding data processing on small or medium sized mobile robots is used in many research projects and starts distributing in more and more applications. This is also the case for agricultural robotics where previously single heavy machines were used. Several recent agricultural robotic systems (in research and in industry) (1) are used outdoors and need to be energy efficient, (2) have to perform complex tasks, e.g., using machine learning (ML) computation like Convolutional Neural Network (CNN), also on high-dimensional multi-spectral image data, and (3) are used in multiple small distributed systems rather than single large vehicles. These properties lead to the contradicting demands for the computation hardware of high computational power and low power consumption and small size. Previous work has shown that using Field Programmable Gate Arrays (FPGA) can lead to a higher energy efficiency compared to standard, e.g., CPU-based ML computation. Furthermore, using a hybrid FPGA with CPU cores and programmable user logic in the same device can lead to a one-chip solution for mobile systems or IoT devices. Presented in this work in progress paper are first results of FPGA-based implementations of ML (CNN) inference on RGB and multi-spectral image data, as part of an agricultural mobile robot research project.