Bayesian networks for obstacle classification in agricultural environments

Autonomous navigation in agricultural environments is a promising research topic for robotics, with several practical applications. This paper presents an obstacle detection system to operate in field scenarios that can accurately discern high and low vegetation from other types of obstacles. Our algorithm is composed by three steps: (i) obstacle detection based on geometric information; (ii) clustering of detected obstacles; and (iii) filtering false positive detections using Bayesian classifiers. Several experimental tests have been carried out in citrus plantations. The results showed that our approach is able to correctly identify obstacles, classifying them as people, bushes, animals, and grass of different heights. In addition, the proposed approach could also be employed as a general framework for stereo-based obstacle detection.