Nowadays, the need to improve the agricultural sector, by increasing efficiency and productivity while reducing losses and pollution, is a key challenge. In this context, the present thesis embrace the broad objective of developing new technologies to serve the agricultural sector, with a specific focus on agricultural robotics. In this thesis, an existing framework called VineSLAM, designed to solve the Simultaneous Localization And Mapping (SLAM) problem specifically designed for robots operating in vineyards and orchards, was first studied in all its tasks. Secondly, a modification related to the extraction of vertical planes (associated with vineyard rows) was implemented. In fact, the original algorithm does not guarantee a realistic characterization of vineyard rows when they do not have a constant orientation along their entire length. The proposed modification involves the extraction of two lines associated with the two vineyard rows, using the RANSAC algorithm twice: the first line is obtained from the entire acquired LiDAR point cloud, while the second is generated from a specific subset of points. These lines are then used to generate two vertical planes. Firstly, the original VineSLAM framework and the modified version were compared in a simulation environment across two different scenarios: a vineyard with straight and parallel rows and a vineyard composed of two sets of four straight, parallel rows, with the second set oriented at 90° with respect to the first. Then, a real-world scenario depicting a curved vineyard with a contour planting configuration was simulated under two conditions, one with canopy coverage and one without canopy coverage, representing respectively summer and winter vineyard scenarios. The estimated trajectories and the generated maps were the objectives of the comparisons using several metrics. Overall, in straight row vineyard scenario, the results suggested that the two algorithm perform approximately the same. In the mixed vineyard, the modified VineSLAM performed similarly to the first scenario, while the original algorithm showed the worst results across all four scenarios. In curved vineyard scenarios, the summer configuration lead to a better performance of the original VineSLAM, whereas the results of the two algorithms in the winter configuration were approximately the same.
Oggigiorno si avverte sempre più la necessità di migliorare il settore dell'agricoltura, rendendolo più efficiente, incrementando la produttività e allo stesso tempo abbattendo sprechi ed inquinamento ad esso associati. In questo contesto si introduce il presente lavoro di tesi con particolare riferimento all'automazione di operazioni tradizionalmente svolte manualmente e, più nello specifico, alla robotica agricola. In questo lavoro di tesi, per prima cosa è stato studiato e descritto in tutte le sue parti un framework esistente, VineSLAM, che permette di risolvere il problema di Simultaneous Localization And Mapping (SLAM), specificamente destinato alla navigazione di robot all'interno di vigneti e frutteti. Successivamente è stata eseguita una modifica al task che riguarda l'estrazione di piani verticali, associati ai filari del vigneto, dalla point cloud acquisita tramite sensore LiDAR. L'algoritmo originale, infatti, non garantisce una corretta caratterizzazione dei filari nel caso in cui questi non risultino rettilinei. La modifica proposta prevede l'estrazione di due rette associate ai due filari del vigneto utilizzando l'algoritmo RANSAC: la prima da tutta la point cloud acquisita, la seconda generata da un insieme di punti accuratamente selezionati. A partire dalle due rette individuate vengono costruiti due piani verticali. Il framework VineSLAM originale e la versione modificata sono stati dapprima confrontati in simulazione in due differenti scenari: un vigneto con filari lineari ed un vigneto con una combinazione di filari orizzontali ed altri disposti a novanta gradi rispetto ad essi. Successivamente, sono state ricreate due simulazioni rappresentanti un possibile caso reale di vigneto con filari curvi in due condizioni stagionali, estiva ed invernale. Le traiettorie e le mappe generate dai due algoritmi sono state confrontate utilizzando specifiche metriche. I risultati mostrano come, nel caso di vigneto con filari rettilinei, i due algoritmi abbiano circa le medesime prestazioni. L'algoritmo modificato presenta gli stessi risultati anche per il caso di vigneto con filari orizzontali e a 90°, mentre l'algoritmo originale presenta le maggiori difficoltà in assoluto. Nel caso di vigneti con filari curvi nella configurazione estiva, l'algoritmo presenta dei risultati migliori rispetto all'algoritmo, mentre evidenti differenze non sono presenti nel caso della configurazione invernale.
A comprehensive study of the VineSLAM framework with a modified vertical plane extraction task
Bongini, Gabriele
2024/2025
Abstract
Nowadays, the need to improve the agricultural sector, by increasing efficiency and productivity while reducing losses and pollution, is a key challenge. In this context, the present thesis embrace the broad objective of developing new technologies to serve the agricultural sector, with a specific focus on agricultural robotics. In this thesis, an existing framework called VineSLAM, designed to solve the Simultaneous Localization And Mapping (SLAM) problem specifically designed for robots operating in vineyards and orchards, was first studied in all its tasks. Secondly, a modification related to the extraction of vertical planes (associated with vineyard rows) was implemented. In fact, the original algorithm does not guarantee a realistic characterization of vineyard rows when they do not have a constant orientation along their entire length. The proposed modification involves the extraction of two lines associated with the two vineyard rows, using the RANSAC algorithm twice: the first line is obtained from the entire acquired LiDAR point cloud, while the second is generated from a specific subset of points. These lines are then used to generate two vertical planes. Firstly, the original VineSLAM framework and the modified version were compared in a simulation environment across two different scenarios: a vineyard with straight and parallel rows and a vineyard composed of two sets of four straight, parallel rows, with the second set oriented at 90° with respect to the first. Then, a real-world scenario depicting a curved vineyard with a contour planting configuration was simulated under two conditions, one with canopy coverage and one without canopy coverage, representing respectively summer and winter vineyard scenarios. The estimated trajectories and the generated maps were the objectives of the comparisons using several metrics. Overall, in straight row vineyard scenario, the results suggested that the two algorithm perform approximately the same. In the mixed vineyard, the modified VineSLAM performed similarly to the first scenario, while the original algorithm showed the worst results across all four scenarios. In curved vineyard scenarios, the summer configuration lead to a better performance of the original VineSLAM, whereas the results of the two algorithms in the winter configuration were approximately the same.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Bongini_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: executive summary tesi
Dimensione
2.35 MB
Formato
Adobe PDF
|
2.35 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Bongini_Tesi.pdf
accessibile in internet per tutti
Descrizione: testo della tesi
Dimensione
16.21 MB
Formato
Adobe PDF
|
16.21 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/253223