A hybrid representation of the environment to improve autonomous navigation of mobile robots in agriculture

Abstract

This paper considers the problem of autonomous navigation in agricultural fields. It proposes a localization and mapping framework based on semantic place classification and key location estimation, which together build a hybrid topological map. This map benefits from generic partitioning of the field, which contains a finite set of well-differentiated workspaces and, through a semantic analysis, it is possible to estimate in a probabilistic way the position (state) of a mobile system in the field. Moreover, this map integrates both metric (key locations) and semantic features (working areas). One of its advantages is that a full and precise map prior to navigation is not necessary. The identification of the key locations and working areas is carried out by a perception system based on 2D LIDAR and RGB cameras. Fusing these data with odometry allows the robot to be located in the topological map. The approach is assessed through off-line data recorded in real conditions in diverse fields during different seasons. It exploits a real-time object detector based on a convolutional neural network called you only look once, version 3, which has been trained to classify a considerable number of crops, including market-garden crops such as broccoli and cabbage, and to identify grapevine trunks. The results show the interest in the approach, which allows (i) obtaining a simple and easy-to-update map, (ii) avoiding the use of artificial landmarks, and thus (iii) improving the autonomy of agricultural robots.