ORCHNet: A Robust Global Feature Aggregation approach for 3D LiDAR-based Place recognition in Orchards

TL;DR

ORCHNet addresses robust place recognition in orchard environments using 3D LiDAR data by fusing multiple global feature aggregations into a single descriptor, trained with a LazyTriplet loss. The architecture combines PointNet or ResNet50-based feature extractors with MAC, SPoC, and GeM pooling, merged via a trainable fusion to form a robust global descriptor. Empirical results on summer and autumn orchard sequences show improved recognition performance and cross-season robustness, and integration into an AMCL localization framework demonstrates practical gains for loop closure. The method is validated on a real-world orchard dataset and the authors provide public code for reproducibility.

Abstract

Robust and reliable place recognition and loop closure detection in agricultural environments is still an open problem. In particular, orchards are a difficult case study due to structural similarity across the entire field. In this work, we address the place recognition problem in orchards resorting to 3D LiDAR data, which is considered a key modality for robustness. Hence, we propose ORCHNet, a deep-learning-based approach that maps 3D-LiDAR scans to global descriptors. Specifically, this work proposes a new global feature aggregation approach, which fuses multiple aggregation methods into a robust global descriptor. ORCHNet is evaluated on real-world data collected in orchards, comprising data from the summer and autumn seasons. To assess the robustness, we compare ORCHNet with state-of-the-art aggregation approaches on data from the same season and across seasons. Moreover, we additionally evaluate the proposed approach as part of a localization framework, where ORCHNet is used as a loop closure detector. The empirical results indicate that, on the place recognition task, ORCHNet outperforms the remaining approaches, and is also more robust across seasons. As for the localization, the edge cases where the path goes through the trees are solved when integrating ORCHNet as a loop detector, showing the potential applicability of the proposed approach in this task. The code will be publicly available at:\url{https://github.com/Cybonic/ORCHNet.git}

Figures (8)

• Figure 1: Representation of the proposed ORCHNet integrated in a localization framework.
• Figure 2: ORCHNet's architecture as part of a retrieval-based place recognition task. ORCHNet receives as input a point cloud $S_i$, which is down-sampled and preprocessed, returning $S'_i$. From $S'_i$, local features $Z$ are extracted, using a feature extractor. The local features are fed into the global feature aggregation module, which fuses the outputs of GeM, SPoC, and MAC into a global descriptor $D$. The global descriptor is used to query the database, which based on a similarity metric, returns the top N loop candidates.
• Figure 3: Training approach using the LazyTriplet loss. The anchor-positive pair has to be from the same physical (i.e. close within the same line), and the anchor-negative pair has to be from distinct places. The LazyTriplet loss is computed by measuring the similarity distance between the anchor-positive and the anchor-negative pair.
• Figure 4: Illustration of the orchards and the robot platform used to collect data. The orchards are located in the southern part of the United Kingdom. The orchards were split in rows, here the autumn sequence is illustrated, which was split into 6 rows.
• Figure 5: Illustration of the points per frame distributions of a) sequence summer and b) sequence autumn.