Semantic Landmark Particle Filter for Robot Localisation in Vineyards

Rajitha de Silva, Jonathan Cox, James R. Heselden, Marija Popović, Cesar Cadena, Riccardo Polvara · 2026

Read Semantic Landmark Particle Filter for Robot Localisation in Vineyards by Rajitha de Silva, Jonathan Cox, James R. Heselden, Marija Popović, Cesar Cadena, Riccardo Polvara (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-10847.txt.

Introduction

Rajitha de Silva and coauthors tackle a hard task in vineyard robotics. Parallel crop rows look so alike to a robot's sensors that it can settle in the wrong row. Their semantic landmark particle filter uses found trunks as row bounds in its model. It also uses satellite positioning as a light prior. In field trials across ten rows, the method cut pose error sharply. It beat geometry-only and vision-based baselines and raised row correctness.

Catalog record

CREATORRajitha de Silva, Jonathan Cox, James R. Heselden, Marija Popović, Cesar Cadena, Riccardo Polvara
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34576
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.10847

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Semantic Landmark Particle Filter for Robot Localisation in Vineyards Rajitha de Silva, Jonathan Cox, James R. Heselden, Marija Popović, Cesar Cadena, Riccardo Polvara Abstract Reliable localisation in vineyards is hindered by row-level perceptual aliasing: parallel crop rows produce nearly identical LiDAR observations, causing geometry-only and vision-based SLAM systems to converge towards incorrect corridors, particularly during headland transitions. We present a Semantic Landmark Particle Filter (SLPF) that integrates trunk and pole landmark detections with 2D LiDAR within a probabilistic localisation framework. Detected trunks are converted into semantic walls , forming structural row boundaries embedded in the measurement model to improve discrimination between adjacent rows. GNSS is incorporated as a lightweight prior that stabilises localisation when semantic observations are sparse. Field experiments in a 10-row vineyard demonstrate consistent improvements over geometry-only (AMCL), vision-based (RTAB-Map), and GNSS baselines. Compared to AMCL, SLPF reduces Absolute Pose Error by 22% and 65% across two traversal directions; relative to a NoisyGNSS baseline, APE decreases by 65% and 61%. Row correctness improves from 0.67 to 0.73, while mean cross-track error decreases from 1.40 m to 1.26 m. These results show that embedding row-level structural semantics within the measurement model enables robust localisation in highly repetitive outdoor agricultural environments. I Introduction Accurate localisation is essential for long-term autonomy in vineyards and orchards, where robots perform monitoring, spraying, and yield estimation. However, vineyards consist of highly repetitive parallel rows that induce severe perceptual aliasing, making adjacent corridors geometrically indistinguishable for geometry-only localisation systems. Seasonal vegetation changes further reduce robustness by introducing transient structure while leaving only trunks and support poles spatially stable. LiDAR-based localisation remains widely used due to its robustness to illumination changes, yet repetitive crop geometry leads to persistent wrong-row hypotheses that standard likelihood models struggle to reject [ 1 , 19 ] . As illustrated in Fig. 1 , adjacent vineyard rows generate nearly identical LiDAR signatures under geometry-only models. Recent work has explored LiDAR-vision fusion to mitigate these limitations. Vision-based trunk detection combined with 2D LiDAR achieves accurate row following in orchards [ 22 ] , while trunk-aware graph-SLAM and keypoint-semantic integration improve loop closure and feature distinctiveness in vineyards [ 20 , 7 ] . These studies demonstrate the value of stable semantic landmarks. However, most fusion approaches focus on reactive navigation or optimisation-based SLAM. They do not explicitly address row aliasing within a probabilistic global localisation framework. In particular, trunk detections are typically treated as isolated landmarks rather than as structural constraints capable of disambiguating parallel corridors. We propose a Semantic Landmark Particle Filter (SLPF) that integrates RGB-D trunk and pole detections with 2D LiDAR observations in a probabilistic formulation. The key idea is to convert stable landmarks into semantic walls : row-level structural boundaries inferred from the spatial arrangement of trunks and support poles. Instead of treating each detection independently, neighbouring landmarks are interpreted as belonging to the same physical row, forming continuous corridor constraints that reflect the underlying vineyard layout. These semantic walls are embedded directly within the particle-filter likelihood, strengthening discrimination between adjacent rows and explicitly penalising cross-row hypotheses under perceptual aliasing. GNSS is included as a lightweight probabilistic prior that primarily stabilises localisation in headlands where semantic observations become sparse. Fig. 1: Vineyard semantics detection (left) projected onto the landmark map (right) for likelihood calculation. Green: vine trunks, Blue: support poles. The contributions of this paper are threefold. First, we introduce semantic walls, transforming sparse trunk and pole detections into row-aligned structural constraints that improve row-level discrimination. Second, we embed these constraints directly within a semantic-LiDAR particle filter, enabling explicit rejection of wrong-row hypotheses in repetitive environments. Third, we incorporate adaptive GNSS weighting based on semantic observation richness, improving robustness during headland transitions while preserving lightweight sensor requirements. Real-world experiments demonstrate improved row correctness and recovery compared to geometry-only and visual-SLAM baselines. Code and dataset available at: https://github.com/LCAS/Outdoor_SLPF . II Related Work Robot localisation in vineyards presents several challenges due to the environment’s repetitive row structure, seasonal appearance changes, and uneven terrain conditions [ 1 , 19 ] . Visual sensors such as LiDAR and cameras are prone to perceptual aliasing in repetitive environments and are sensitive to lighting variation [ 12 ] . Non-visual sensors, including GNSS and ultrasonic systems, can be affected by canopy occlusion, multipath, and signal degradation [ 5 , 6 ] . Robust localisation remains essential for enabling long-term autonomy and digital twin applications in vineyard management [ 21 , 15 ] . LiDAR-based particle filters such as Adaptive Monte Carlo Localization (AMCL) [ 8 ] have been widely applied in structured environments and extended to agricultural domains [ 2 , 17 ] . While their probabilistic formulation allows tolerance to map-sensor mismatch [ 3 ] , their measurement models rely purely on geometric consistency. In vineyards, where parallel rows generate near-identical range signatures, geometry-only likelihoods may maintain persistent wrong-row hypotheses, particularly during headland transitions. Visual SLAM approaches have also been explored in vineyards [ 14 , 20 ] . Papadimitriou et al. [ 20 ] incorporate trunk detections into a graph-SLAM formulation, demonstrating improved loop closure robustness compared to non-semantic baselines. Similarly, De Silva et al. [ 7 ] introduce keypoint-semantic integration to improve visual feature distinctiveness in repetitive agricultural environments. These works highlight the value of semantically stable landmarks such as vine trunks, yet they remain embedded within optimisation-based SLAM back-ends rather than explicitly reshaping the measurement likelihood in a global probabilistic filter. Several recent works explore LiDAR-vision fusion for row navigation. For example, Shi et al. [ 22 ] combine YOLO-based trunk detection with 2D LiDAR via ray projection to extract navigation lines in dense orchard environments, achieving high lateral accuracy. Such approaches effectively support corridor following but assume locally continuous vegetation and are primarily designed for reactive control rather than maintaining a global pose posterior. Stable-point segmentation methods further distinguish permanent structural elements from seasonal vegetation to enhance long-term consistency in 3D LiDAR mapping [ 11 ] . While improving robustness to appearance changes, these approaches typically operate within dense SLAM frameworks and do not directly address row-level aliasing in global localisation. GNSS-based localisation remains common in agricultural robotics [ 18 ] . Sensor-fusion approaches combining SLAM and GNSS dynamically adjust weighting based on signal quality metrics [ 24 ] . While effective in GNSS-degraded scenarios, such methods often assume RTK precision or rely on factor-graph optimisation. In contrast to prior work, our approach makes three conceptual shifts. First, rather than treating trunk detections as isolated landmarks, we connect adjacent stable landmarks into piecewise-linear semantic walls, forming row-aligned structural constraints that persist even when individual plants are missing. Second, these semantic walls are embedded directly within the particle-filter measurement likelihood, enabling explicit penalisation of cross-row hypotheses under perceptual aliasing. Third, GNSS is incorporated as an adaptive probabilistic prior whose influence depends on semantic observation richness, providing global anchoring without dominating the estimation process. Together, these design choices specifically address the row-aliasing failure mode characteristic of repetitive agricultural environments. More broadly, they illustrate how semantically stable, spatially organised landmarks can be converted into structural constraints directly within a probabilistic measurement model. While demonstrated in vineyards, this formulation is applicable to other structured outdoor domains exhibiting parallel or repetitive geometry, such as

Read the complete text