In recent years, robust and accurate localization of autonomous agricultural vehicles has become a key challenge in the development of intelligent systems for precision farming. This thesis presents two LiDAR-based odometry methods, KISS-ICP and Kinematic-ICP, implemented to estimate the trajectory of an autonomous tractor operating in vineyard environment by iteratively aligning LiDAR scans with an incrementally built local map. KISS-ICP uses a constant velocity motion model as initial guess and optimizes point-to-point correspondences without applying kinematic constraints, producing corrections in SE(3). Kinematic-ICP, instead, relies on an external odometry source to provide an initial guess to the registration algorithm and performs constrained optimization in SE(2), projecting the result into SE(3) using a unicycle model. Both algorithms were validated offline using real-world data collected in vineyards, by comparing the estimated trajectories with GNSS data, used as ground truth. KISS-ICP was analyzed in terms of sensitivity to LiDAR information content (through ground removal, region of interest and their combination) and spatial resolution. Kinematic-ICP was evaluated by varying the balance between odometry and LiDAR contributions. The results show that both methods achieve reliable performance in terms of accuracy and robustness. KISS-ICP provides higher localization accuracy but generates irregular trajectories that are not physically feasible for a wheeled vehicle. Instead, Kinematic-ICP, while slightly less accurate in position, produces smoother and more realistic trajectories, which are consistent with the actual motion of a tractor.
Negli ultimi anni, la localizzazione robusta e accurata di veicoli agricoli autonomi ha assunto un ruolo centrale nello sviluppo di sistemi intelligenti per l’agricoltura di precisione. In questo elaborato vengono presentati due metodi di odometria LiDAR, KISS-ICP e Kinematic-ICP, implementati per stimare la traiettoria di un trattore autonomo operante in vigneto attraverso l’allineamento iterativo delle scansioni LiDAR con una mappa locale costruita in modo incrementale. KISS-ICP utilizza come stima iniziale un modello a velocità costante e ottimizza le corrispondenze punto-punto generando correzioni in SE(3) senza vincoli cinematici. Kinematic-ICP, invece, si affida ad una fonte odometrica esterna per fornire una stima iniziale all'algoritmo di registrazione ed esegue una correzione vincolata in SE(2), successivamente proiettata in SE(3) tramite un modello unicycle. I due algoritmi sono stati validati offline utilizzando dati reali raccolti in vigneto, confrontando la traiettoria stimata con i dati GNSS, usati come riferimento. Per KISS-ICP è stata condotta un’analisi di sensibilità rispetto al contenuto informativo della point cloud (tramite rimozione del terreno, regione di interesse e la loro combinazione) e alla risoluzione spaziale. Per Kinematic-ICP è stata invece analizzata la robustezza rispetto al bilanciamento tra stima odometrica e correzione LiDAR. I risultati mostrano che entrambi i metodi garantiscono prestazioni soddisfacenti in termini di accuratezza e robustezza. KISS-ICP risulta più preciso nella localizzazione, ma produce traiettorie irregolari e non compatibili con la dinamica reale di un trattore. Invece, Kinematic-ICP, sebbene leggermente meno accurato a livello di posizione, genera traiettorie più fluide, regolari e fisicamente realizzabili, in linea con il comportamento atteso del veicolo.
LiDAR-based localization for an autonomous vineyard tractor using point-to-point ICP algorithms
Pantella, Edoardo
2024/2025
Abstract
In recent years, robust and accurate localization of autonomous agricultural vehicles has become a key challenge in the development of intelligent systems for precision farming. This thesis presents two LiDAR-based odometry methods, KISS-ICP and Kinematic-ICP, implemented to estimate the trajectory of an autonomous tractor operating in vineyard environment by iteratively aligning LiDAR scans with an incrementally built local map. KISS-ICP uses a constant velocity motion model as initial guess and optimizes point-to-point correspondences without applying kinematic constraints, producing corrections in SE(3). Kinematic-ICP, instead, relies on an external odometry source to provide an initial guess to the registration algorithm and performs constrained optimization in SE(2), projecting the result into SE(3) using a unicycle model. Both algorithms were validated offline using real-world data collected in vineyards, by comparing the estimated trajectories with GNSS data, used as ground truth. KISS-ICP was analyzed in terms of sensitivity to LiDAR information content (through ground removal, region of interest and their combination) and spatial resolution. Kinematic-ICP was evaluated by varying the balance between odometry and LiDAR contributions. The results show that both methods achieve reliable performance in terms of accuracy and robustness. KISS-ICP provides higher localization accuracy but generates irregular trajectories that are not physically feasible for a wheeled vehicle. Instead, Kinematic-ICP, while slightly less accurate in position, produces smoother and more realistic trajectories, which are consistent with the actual motion of a tractor.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_Pantella_Tesi.pdf
non accessibile
Descrizione: testo tesi
Dimensione
30.37 MB
Formato
Adobe PDF
|
30.37 MB | Adobe PDF | Visualizza/Apri |
|
2025_07_Pantella_Executive Summary.pdf
non accessibile
Descrizione: executive summary
Dimensione
8.46 MB
Formato
Adobe PDF
|
8.46 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/10589/239836