Semantic Landmark Particle Filter for Robot Localisation in Vineyards
Authors: Rajitha de Silva, Jonathan Cox, James R. Heselden, Marija Popović, Cesar Cadena, Riccardo Polvara
Organizations: Lincoln Centre for Autonomous Systems (L-CAS), University of Lincoln, UK. · MAVLab, TU Delft, Netherlands. · Robotics Systems Lab, ETH Zurich, Switzerland.
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.
Figures & tables
Fig. 1: Vineyard semantics detection (left) projected onto the landmark map (right) for likelihood calculation. Green: vine trunks, Blue: support poles.
Fig. 2: Vineyard structural map with RTK-GNSS ground-truth traverses for the two experiments. Orange circles and green squares indicate row posts and vines, while orange and green line segments denote pole- and trunk-based semantic walls. Red (Experiment 1) and blue (Experiment 2) trajectories are the RTK paths; start/end markers indicate traversal direction.
Fig. 3: Overview of the Semantic Particle Filter. Instance masks from segmentation are combined with depth to generate BEV projections of vineyard landmarks. These projections label near-field LiDAR returns as semantic observations for the particle filter, augmented with a NoisyGNSS prior.
Fig. 4: Raw Absolute Pose Error (APE) trajectories comparing SLPF (ours), AMCL+NoisyGNSS (Kalman fusion), RTAB-Map RGBD, and NoisyGNSS (subsampled to facilitate visualisation). The dashed black line is ground truth; coloured paths show method deviations.
Fig. 5: Zoomed comparison during a headland transition (Experiment 1). AMCL+NoisyGNSS remains aligned to an adjacent row despite locally smooth tracking, while SLPF recovers through semantic wall constraints and adaptive GNSS weighting.
Variant
APE ↓
RPE 2m ↓
XT mean ↓
Row corr. ↑
Switches ↓
SLPF (full map)
1.00 ± 0.03
3.33 ± 0.02
1.31 ± 0.05
0.72 ± 0.02
33.00 ± 0.82
DetDrop 20%
1.02 ± 0.08
3.37 ± 0.06
1.31 ± 0.05
0.73 ± 0.01
33.33 ± 4.11
DetDrop 40%
1.05 ± 0.09
3.33 ± 0.04
1.26 ± 0.05
0.74 ± 0.03
30.67 ± 0.94
MapRemove 30%
1.05 ± 0.04
3.37 ± 0.03
1.27 ± 0.02
0.72 ± 0.01
30.33 ± 2.05
MapRemove 50%
1.13 ± 0.13
3.36 ± 0.06
1.32 ± 0.02
0.72 ± 0.01
31.00 ± 0.82
TABLE II: Experiment 1 robustness under Detection Dropout and Section Landmark Removal. Values are mean ± std over three runs.
University Institute for Engineering Research, Miguel Hernández University, Avda. de la Universidad s/n, Edificio Innova, Elche, 03202, Alicante, Spain