An Anomaly Detection Approach to Monitor the Structured-Based Navigation in Agricultural Robotics

Abstract

Local perception navigation methods allow agricultural robots to accurately track crop row structures while performing automated farming tasks. The integration of these methods as a part of a fully autonomous navigation solution requires continuous assessment of their reliability since they rely solely on sensor data in a changing and unpredictable environment. This paper presents a data-driven monitoring approach for the task of structure-based navigation in agriculture. The proposed method applies semi-supervised anomaly detection, aiming to learn a model of normal scene geometry that characterizes a domain of reliable execution of the considered task. To this end, a convolutional neural network was trained in one-class classification fashion on Hough representations of LiDAR point clouds. In experimentation, the learned normal model was used to derive a confidence measure for a LiDAR-based tracking algorithm allowing its integration as a part of a hybrid navigation solution in vineyards for a commercial robotic platform.