Field scene recognition for self-localization of autonomous agricultural vehicle

In this study, a field scene recognition system was developed to estimate a self-position of a traveling vehicle along a farm road by using an original capture system with three cameras, a vector quantization method to express the features of field scenes, a machine learning based scene recognition algorithm, and a vehicle position estimation algorithm with an original voting method. The potential of our system was demonstrated through five experiments performed over four months. In the experiments, the system could robustly estimate the vehicle position with the accuracy less than 1 m at the processing speed of approximately 2.0 Hz when the vehicle was driven straight along a traveling line on the targeted two types of roads: a surfaced road and an unsurfaced road, at the driving speed of 0.5 m/s. The results demonstrated an applicability of our system to navigate an autonomous agricultural robot vehicle without using GNSS.