Fusão de sensores na autolocalização de agrobots em ambientes internos: Uso combinado do algorítmo YOLO e dados de RSSI

Abstract

This work evaluates the effect of the data fusion method (sensors) on the performance of self localization procedures for mobile agrobots inside greenhouses. To do so, it compares the estimation errors of coordinates and the processing time of two self-localization procedures. The first procedure, CVAutoL, utilizes information extracted from images via computer vision techniques. The second, FusionAutoL, fuses data extracted from images with RSSI data. The detection of the reference object for position estimation in the images was performed using the YOLOv8 algorithm. Both procedures employed a nonlinear regression routine, based on a support vector machine, to fit a function that relates the attributes to the agrobot’s coordinates. The errors in the distance between the estimated coordinates and the actual agrobot coordinates (E2) and the difference between the estimated and actual distance from the agrobot to the marker (EO) were compared. Three versions of YOLOv8-Seg were used during the tests, nano, medium, and extra-large to verify whether data fusion allows the use of shallow networks without compromising the precision of self-localization. The results obtained indicate that the data fusion procedure resulted in lower errors, with E2 and EO. They also suggest that the data fusion enabled the use of shallow and faster versions of YOLO without increasing the error in position estimates.