Classification of scenery using multinomial logistic regression in a sugarcane crop

• Leonardo Bonacini USP
• Handel Emanuel Natividade Peres USP
• Vitor Akihiro Higuti EarthSense Co.
• Vivian Suzano Medeiros USP
• Marcelo Becker USP
• Mário Luiz Tronco USP

Resumo

Autonomous navigation is an important skill for robots that perform tasks in the agricultural field. The main sensors used for this application are GPS, camera, and LiDAR. Navigation methodologies for such robots either focus on global localization and path planning based on GPS information or local path planning employing a perception system, using cameras and LiDAR sensors to extract crop characteristics to suggest control actions. Since mobile platforms are limited in both energetic and computational power, this situation highlights the need for a high-level system that chooses which navigation algorithm should be employed depending on the environment. Based on this information, this work presents a system for classifying agricultural scenarios based on LiDAR data. In this work, we separate four conditions concerning the position of the crop around the robot: ‘betweenRowCrop’, ‘leftRowCrop’, ‘rightRowCrop’, and ‘noRowCrop’. We used two Hokuyo UTM30-LX LiDARs on an agricultural robot to collect extensive data on a sugarcane crop. With the data from the LiDAR, several statistical measures are obtained and, for each variable, outlier samples are removed using the IQR rule. After that, these variables are used as predictors in a Multinomial Logistic Regression to classify agricultural scenes. The final model presented an accuracy of approximately 99%. This indicates that this model could be a promising solution for classifying the agricultural scenery the robot encounters and then passing this information to the robot's navigation system to choose the appropriate navigation method.

Palavras-chave: Laser radar, Navigation, Robot vision systems, Crops, Sensor phenomena and characterization, Cameras, Path planning