Semantic-Weighted Adaptive Particle Filter

• SWA-PF is a semantic-guided particle filtering approach that uses semantic consistency instead of raw geometric matching to weight particles for localization.
• It employs adaptive techniques such as semantic-aligned initialization, clustering, and covariance monitoring to optimize particle distribution and convergence.
• Empirical evaluations demonstrate that SWA-PF delivers significant runtime improvements and robust meter-scale localization in challenging, GNSS-denied scenarios.

Semantic-Weighted Adaptive Particle Filter (SWA-PF) denotes a semantic-cued particle-filtering approach in which particle importance weights are determined by semantic consistency rather than by raw geometric or appearance matching. In its explicit formulation for GNSS-denied UAV localization, SWA-PF combines semantic segmentation from UAV imagery and satellite imagery with an optimized adaptive particle filter that estimates a 4-DoF pose using a low-resolution 2D satellite map (Yuan et al., 17 Sep 2025). A closely related precursor appears in underwater robotics, where a “semantic-aided particle filter” for AUV self-localization modifies the measurement model of a standard SIR particle filter so that particles are weighted by semantic object matches and mismatch penalties instead of purely geometric measurements (Maurelli et al., 2019). Taken together, these works define SWA-PF as a semantic weighting strategy embedded in Monte Carlo localization, with adaptation arising from likelihood shaping, particle initialization, and convergence detection rather than from a distinct Bayesian filtering formalism.

1. Terminological scope and historical relation

The term SWA-PF is explicitly introduced in “SWA-PF: Semantic-Weighted Adaptive Particle Filter for Memory-Efficient 4-DoF UAV Localization in GNSS-Denied Environments” (Yuan et al., 17 Sep 2025). That work presents SWA-PF as a response to the limitations of retrieval-based and feature-heavy cross-view localization systems, particularly their “suboptimal real-time performance, environmental sensitivity, and limited generalization capability, particularly in dynamic or temporally varying environments” (Yuan et al., 17 Sep 2025).

A semantically analogous construction had already appeared in “A semantic-aided particle filter approach for AUV localization” (Maurelli et al., 2019). That paper does not explicitly use the acronym SWA-PF and does not present a separate Bayesian filter family. Instead, it extends a standard SIR particle filter by replacing a geometric likelihood with a semantic one. The correspondence is direct at the mechanism level: particle weights depend on semantic object matches and penalties for unmatched entities, while resampling propagates particles according to those semantic weights (Maurelli et al., 2019).

Work	Domain	Core characterization
(Maurelli et al., 2019)	AUV localization	Semantic refinement of a standard SIR particle filter
(Yuan et al., 17 Sep 2025)	UAV localization	Explicit “Semantic-Weighted Adaptive Particle Filter” for 4-DoF pose estimation

This relation is significant because it separates nomenclature from mechanism. The 2019 AUV work establishes semantic weighting inside a particle filter for underwater localization, whereas the 2025 UAV work formalizes that pattern under the SWA-PF name and adds adaptive initialization, clustering, and GPU-oriented optimizations. A plausible implication is that SWA-PF is best understood as a design pattern within particle filtering rather than as a wholly separate inferential paradigm.

2. Problem formulation and state-space representation

In both formulations, the localization problem is posed in environments where external global positioning is unavailable or inadequate. For the AUV case, GPS is unavailable underwater, so the vehicle must estimate its pose from onboard sensing alone; the authors position semantic information as an alternative to acoustic or geometric cues that require expensive measurement simulation (Maurelli et al., 2019). For the UAV case, the target setting is GNSS-denied localization under cross-view, altitude-varying, and temporally changing conditions (Yuan et al., 17 Sep 2025).

The AUV formulation starts from a standard particle-filter approximation of the Bayes filter. The belief over state is represented by samples as

$p(x) \approx \dfrac{1}{N} \sum\limits_{i} \delta_{x^{i}(x)$

and the posterior after incorporating the observation history is approximated as

$$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$

with importance weights

$$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$

The filter is described through the usual SIR stages: sampling, importance weighting, and resampling. The paper states that the state estimate is updated using the AUV commands $u_t$ and observations $z_t$, but it does not specify a detailed motion model $p(x_t|x_{t-1},u_t)$ or a closed-form transition equation (Maurelli et al., 2019).

The UAV SWA-PF formulation is explicitly 4-DoF. The state is defined as

$$x_{t} = \begin{bmatrix} x & y & h & \theta \end{bmatrix},$$

where $x$, $y$, and $h$ are the 3D positional coordinates and $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$0 is the yaw angle (Yuan et al., 17 Sep 2025). The motivation is that satellite images are inherently 2D, whereas UAV localization must account for altitude and heading; accordingly, the particle distribution is expanded from a 2D plane into 3D space plus yaw. The Bayesian recursion is written as

$$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$1

$$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$2

Its state transition uses a kinematic update driven by control input $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$3, representing odometry-derived rigid-body motion from the previous frame. The paper writes

$$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$4

with noise $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$5 applied to $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$6, $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$7, $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$8, and $$p(x_t | z_{0:t}) \propto \sum\limits_{i} w^{(i)} \delta_{x_{t}^{(i)}}$$9 (Yuan et al., 17 Sep 2025).

3. Semantic observation models

The defining shift in SWA-PF is from direct geometric or appearance matching to semantic observation modeling. In the AUV formulation, the authors frame this explicitly as moving from “signal to symbol” (Maurelli et al., 2019). Rather than simulating dense sensor returns from surfaces, the filter reasons over symbolic environmental objects such as man-made structures stored in a semantic map.

An observation from the robot is represented as

$$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$0

and the robot-generated predicted observation for a particle is

$$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$1

Here $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$2 and $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$3 are the relative range and bearing of each object with respect to the robot pose (Maurelli et al., 2019). The semantic classes are not a long ontology; the paper emphasizes “objects” and “man-made structures” in the environment. The object-recognition step that converts sensor signals to symbols is assumed to exist upstream and is outside the paper’s scope.

In the UAV SWA-PF formulation, the observation $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$4 is a semantic UAV top-view image represented as $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$5, where $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$6 and $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$7 are the image coordinates and $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$8 is the semantic label (Yuan et al., 17 Sep 2025). For a hypothesized particle pose, the algorithm queries the semantic map $$w^{(i)} = \dfrac{p(x_t | z_{1:t}, u_{0:t})}{\pi(x_t | z_{1:t}, u_{0:t})}.$$9 at that location and compares semantic consistency. The semantic map is derived from segmentation on both the UAV view and the satellite view.

The segmentation pipelines are asymmetric. For satellite imagery, the authors adapt a VGG-pretrained U-Net, crop images to $u_t$0 px, and use augmentation including random scaling, rotation, cropping, and flipping (Yuan et al., 17 Sep 2025). The paper states that the model is trained on only three annotated satellite images from SemanticMAFS and still achieves “satisfactory segmentation performance” because the scenes are dense and share a similar urban context. For UAV top-view imagery, an initial U-Net-based attempt with only four classes produced misclassifications, especially rooftops being labeled as ground. The final model is a SegFormer-B0 trained with the semantic classes roof, wall, ground, road, vegetation, water body, and vehicle; the satellite semantic classes are building, ground, road, and vegetation. The SegFormer input size is $u_t$1, the backbone is ImageNet-pretrained, and training proceeds in two stages on a single RTX 4060Ti: 50 epochs with the backbone frozen at batch size 32, then 200 epochs fine-tuning with batch size 8, using AdamW with learning rate $u_t$2, weight decay $u_t$3, and cosine annealing. The reported segmentation performance is $u_t$4 and pixel accuracy $u_t$5 (Yuan et al., 17 Sep 2025).

4. Semantic weighting and adaptive likelihood shaping

In the AUV semantic-aided particle filter, the importance weighting step is modified directly by semantic observations. The weight is computed by comparing the observed semantic set $u_t$6 and the predicted semantic set $u_t$7 (Maurelli et al., 2019). For the $u_t$8 objects common to both sets, the paper states that a mixture of two families of Gaussians is used, centered at zero and evaluated at residuals

$u_t$9

The calculated value is then adjusted for the $z_t$0 objects that appear in one observation but not the other, where

$z_t$1

Agreement on common semantic entities increases particle weight, while missing or extra entities penalize it. The paper does not provide a more explicit normalized update equation, but the standard interpretation is that these semantic likelihood values become the particle weights before normalization and resampling (Maurelli et al., 2019).

In the explicit UAV SWA-PF, the semantic weighting mechanism is formulated through a Semantic Weighted Distance Map (SWDM) (Yuan et al., 17 Sep 2025). For each semantic label, the algorithm computes the distance from each pixel to the nearest pixel with the same semantic class:

$z_t$2

This is described as calculating Euclidean distances to the nearest semantically equivalent point across all pixels. A center distance field matrix $z_t$3 is then introduced to encourage center alignment and reduce the influence of ambiguous borders. The particle weight is given as

$z_t$4

The notation is awkwardly formatted in the paper, but its stated semantics are specific: $z_t$5 is a semantic-class-specific coefficient, $z_t$6 encodes semantic distance, $z_t$7 biases the center region, and $z_t$8 is a normalization constant that introduces slower convergence (Yuan et al., 17 Sep 2025). After weight computation, the weights are normalized so that the sum over all particles is 1.

The two weighting schemes share a common principle. In both, semantic agreement drives the likelihood, and mismatch penalties prevent spurious consistency. The difference is representational: the AUV work compares sets of symbolic objects with relative pose parameters, whereas the UAV work compares dense semantic fields through nearest-semantic distances and a center-biased correction. This suggests two operational regimes of SWA-PF-like filtering: object-level semantic matching and semantic-field distance matching.

5. Adaptation, convergence control, and computational optimization

The adaptive dimension of SWA-PF is most explicit in the UAV formulation (Yuan et al., 17 Sep 2025). The paper identifies particle transformation cost—especially rotation and translation for large particle sets—as a major bottleneck. To reduce that cost, UAV-view image rotation is discretized into 100 steps of $z_t$9 each, and each particle is matched to the nearest yaw bin. The implementation also simplifies matrix operations into two-stage multiplication and accelerates batch processing using optimized matrix stacking on GPU. With these optimizations, the paper reports processing 20,000 particle matrices in approximately 4.9 seconds.

Pose estimation is also made adaptive at the posterior level. Rather than averaging all particles, the method applies DBSCAN to remove outliers, clusters the remaining particles, computes cluster centroids, and monitors the covariance of the posterior particle set. When covariance drops below a threshold, the cluster centroid is taken as the final position estimate and projected onto the global map (Yuan et al., 17 Sep 2025). The adaptation therefore lies in convergence detection from the evolving particle distribution rather than in a fixed iteration count.

Initialization is another adaptive component. The paper compares full-space randomization, layered randomization, and center semantic class alignment. The best strategy is semantic-aligned initialization, in which particles are clustered around semantically salient central regions in the UAV image; layered randomization is faster than full-space randomization but degrades when actual altitude does not match preset layers (Yuan et al., 17 Sep 2025).

The AUV precursor does not introduce such an elaborate adaptive architecture, but it does establish the central computational premise of semantic weighting: generating a semantic observation from a semantic map is much cheaper than simulating geometric sensor returns from surrounding structures (Maurelli et al., 2019). In the geometric version, the filter must simulate distances to surrounding structures, typically requiring ray tracing or expensive geometric calculations. In the semantic version, the predicted observation is a set of symbolic objects with relative positions, so the filter avoids those costly operations. The paper notes that even its geometric baseline already used analytic formulas rather than full ray tracing, which implies that the real-world computational advantage of the semantic method could be even larger.

6. Evaluation, empirical behavior, and limitations

The AUV study is entirely simulated and was integrated into a ROS-based AUV simulator at Jacobs University Bremen built on the Morse simulator (Maurelli et al., 2019). The environment is a $p(x_t|x_{t-1},u_t)$0 m area with the AUV moving near various man-made structures and following a planned trajectory. The comparison is against a standard geometric approach. Evaluation uses 100 runs per approach, with average particle variance and average localization error over time. Execution time is reported as average time in nanoseconds over 20 runs: semantic-aided localization takes 1,171,934,146 ns on average, while the geometric approach takes 6,977,878,010 ns, summarized as roughly a 1:6 speed ratio. The semantic method shows slightly better localization variance and error curves, but the paper states that the accuracy improvement is modest compared with the speed gain (Maurelli et al., 2019).

The UAV SWA-PF evaluation is centered on the MAFS dataset, a large-scale Multi-Altitude Flight Segments dataset built from UAV footage over 14 university campuses in Hangzhou, China, plus one larger $p(x_t|x_{t-1},u_t)$1 area labeled MAFS-14 (Yuan et al., 17 Sep 2025). Fixed-altitude flights are conducted at 100 m, 200 m, 300 m, 400 m, and 500 m, while variable-altitude trajectories span 150 m to 500 m. Each flight is recorded in 4K at 30 fps using a DJI MAVIC III E and synchronized with onboard IMU data, giving per-frame latitude, longitude, altitude, yaw, and velocity. A corresponding semantic annotation set, SemanticMAFS, labels UAV images into seven classes and satellite imagery into four classes.

Experiments are conducted on MAFS-03 and MAFS-10 as fixed-altitude 200 m scenarios, and on MAFS-14 as the difficult variable-altitude scenario. Each trajectory is tested 20 times, and both best and average performance are reported (Yuan et al., 17 Sep 2025). The baseline particle settings are 5,000 particles for the fixed-altitude cases and 40,000 particles for MAFS-14. The comparisons include MI, HIST, SIFT, AKAZE, BRISK, ORB, SSPT, DenseUAV, and OS-FPI. On MAFS-10, SWA-PF achieves fitting time of 7 s and finish time of 25 s, whereas SIFT takes 1630 s fitting and 3338 s total, MI takes 433 s fitting and 1301 s total, HIST takes 143 s fitting and 524 s total, BRISK takes 1191 s fitting and 3028 s total, and ORB takes 90 s fitting and 299 s total. In accuracy terms, SWA-PF reports RMSE 6.5685 m and median error 6.653 m, compared with HIST at 14.317 m RMSE, BRISK at 16.469 m, MI at 21.775 m, and SIFT at 62.825 m. Recall@10 is 97.368% for SWA-PF, while ORB reports 7.894%, BRISK 6.578%, HIST 9.210%, MI 5.263%, and SIFT 0. The deep-learning baselines in this evaluation are also less favorable: SSPT has RMSE 65.958 m, DenseUAV 359.416 m, and OS-FPI 90.114 m, with very low Recall@10 (Yuan et al., 17 Sep 2025).

For the variable-altitude MAFS-14 case, the authors state that the three-dimensional search space is about three orders of magnitude larger than conventional approaches, yet the optimized weighting and altitude-aware particle filter keep convergence practical (Yuan et al., 17 Sep 2025). The paper states that the method localizes within seconds and maintains global positioning error below 10 m overall; another passage characterizes the comparison result as “under 9 meters.” Table III reports error below 9 m on a $p(x_t|x_{t-1},u_t)$2 m map and a normalized error-to-map ratio that is 3.4× and 17.7× lower than two prior methods. The complete pipeline is reported to run in full real time on an Intel i5-12400F CPU and RTX 4060Ti GPU.

The limitations are also explicit. In the AUV case, object recognition may itself be computationally expensive and is not counted in localization timing unless it is already part of the robot’s perception pipeline; the semantic map is deterministic in this work, and future work is directed toward probabilistic semantic maps and real-data validation (Maurelli et al., 2019). In the UAV case, the method struggles when the field of view is dominated by low-distinctiveness regions such as large buildings or vegetation, because there are not enough reference features for spatial alignment. It is also sensitive to perspective disparity between oblique satellite views and UAV top-view images, which can introduce geometric mismatch and larger localization errors. The authors state that if the scene has too few semantic cues, accurate pose estimation becomes difficult. They argue that the semantic-weighted framework is more resistant to seasonal or temporal map changes than raw appearance matching, and that low-resolution satellite experiments show little degradation when map resolution is halved, but they also acknowledge that dynamic environments and temporally varying conditions remain challenging (Yuan et al., 17 Sep 2025).

7. Position within semantic localization research

SWA-PF occupies a specific position at the intersection of semantic mapping, Monte Carlo localization, and cross-view registration. Relative to purely geometric particle filters, the semantic layer modifies the measurement model rather than the state transition model (Maurelli et al., 2019). Relative to retrieval-based or feature-heavy cross-view localization pipelines, the explicit UAV SWA-PF replaces brittle image pairing and dense feature extraction with semantic distance weighting inside a particle filter (Yuan et al., 17 Sep 2025).

This yields a coherent interpretation of the term. The “semantic-weighted” component refers to the observation likelihood: matched semantic entities or semantically proximate regions increase particle weight, while mismatches and ambiguous borders reduce it. The “adaptive” component refers to mechanisms such as mismatch penalties, semantic-aligned initialization, covariance-based convergence detection, and clustering-based posterior consolidation. The “particle filter” component remains standard Bayesian Monte Carlo localization, with sampling, weighting, normalization, and resampling.

Across both the underwater and aerial formulations, the principal significance is computational and representational. Semantic maps permit cheaper observation generation than dense geometric simulation in the AUV case, and semantic distance fields provide robustness to viewpoint, lighting, and seasonal change in the UAV case (Maurelli et al., 2019, Yuan et al., 17 Sep 2025). The empirical record provided by these papers is consistent with that interpretation: the AUV study reports a roughly sixfold timing advantage with modest accuracy gains, while the UAV study reports substantial runtime reductions relative to classical baselines together with meter-scale localization accuracy in GNSS-denied conditions. A plausible implication is that SWA-PF is most effective when a semantic map already exists or can be produced reliably, and when environment structure is sufficiently distinctive for semantic agreement to discriminate between hypotheses.