| Reliable localization in agricultural environments, such as vineyards, remains a significant challenge for autonomous robots due to high structural repetitiveness and the lack of distinctive landmarks. To address these limitations, we present MTL-MinkUNeXt-VINE, an unified multi-task learning architecture capable of performing simultaneous LiDAR Place Recognition and Semantic Place Classification. The system produces a global descriptor for retrieval and efficient classification of the scene into intra-row segments or turnpoints via a dedicated auxiliary branch that extends from the unmodified encoder of the original method. To accurately address the class imbalance in extensive vineyard environments, we use a weighted cross-entropy loss function. Additionally, we apply a median frequency balancing protocol to weight each label. The experimental results indicate that the accuracy of the additional Semantic Place Classification task reaches 95%. The code of MTL-MinkUNeXt-VINE is publicly available for reproduction. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.