Addressing the challenges of loop detection in agricultural environments

Abstract

While visual SLAM systems are well studied and achieve impressive results in indoor and urban settings, natural, outdoor and open-field environments are much less explored and still present relevant research challenges. Visual navigation and local mapping have shown a relatively good performance in open-field environments. However, globally consistent mapping and long-term localization still depend on the robustness of loop detection and closure, for which the literature is scarce. In this work we propose a novel method to pave the way towards robust loop detection in open fields, particularly in agricultural settings, based on local feature search and stereo geometric refinement, with a final stage of relative pose estimation. Our method consistently achieves good loop detections, with a median error of 15cm. We aim to characterize open fields as a novel environment for loop detection, understanding the limitations and problems that arise when dealing with them.

Figures (9)

• Figure 1: Overview of our loop detection pipeline, where $k$ are keypoints for the query image, $d$ is a descriptor of the left query image, $s$ are the similarity scores of the bow search, $I$ are the island sets, and $m$ are the keypoints for the best matching image.
• Figure 2: Planar projection of the GPS-RTK trajectories for the three FieldSAFE sessions.
• Figure 3: Resulting loop detections overlaid on the trajectories for the FieldSAFE dataset sessions. Crosses stand for detected loops, and their colors represent the absolute distance error between the estimated and the ground truth relative pose.
• Figure 4: Image \ref{['sfig:sl_error_boxplots']} shows the distribution (as boxplots) of the absolute error between distances predicted for the loop detections by our method and actual distance given by each dataset's ground-truth. Image \ref{['sfig:sl_distance_boxplots']} shows the distribution of the distance to the loop from each dataset's ground-truth. The number on top of each boxplot shows how many samples are taken into account on computing each boxplot (equal to the number of loops detected in each sequence).
• Figure 5: One example for each session of correct loop detections from our stereo matching system, in a two by two tile to accomodate both stereo images, with lines connecting the keypoints used for the final relative pose estimation. Note that the right stereo images are grayscale, as that is what the camera used in the dataset provides.