Dynamic Points Removal Algorithms

• Dynamic points removal is a set of algorithms designed to identify and eliminate dynamic objects from point clouds to preserve accurate static representations.
• These methods utilize geometric, probabilistic, and learning-based strategies to overcome challenges like occlusion, sensor noise, and computational demands.
• By incorporating multi-resolution analysis and adaptive thresholds, dynamic points removal significantly improves localization, mapping, and downstream perception in robotics.

Dynamic points removal is the class of algorithms and methods for detecting and eliminating observations of non-static, moving, or temporally inconsistent entities from a spatial dataset—typically, point clouds in robotics, SLAM, mapping, or combinatorial optimization. The goal is to distinguish and suppress points originating from dynamic objects (e.g., vehicles, pedestrians, machinery, etc.) so that downstream tasks such as localization, mapping, or optimization can operate on representations that more accurately reflect the static environment. The challenge is that naive accumulation or integration without dynamic removal yields “ghosts”, artifacts, and severe localization errors, especially in dense urban or cluttered indoor environments.

1. Principles and Challenges of Dynamic Points Removal

Dynamic points removal is inherently a spatiotemporal recognition problem that leverages the structural, statistical, or physical priors of static versus dynamic entities. Key challenges include:

• Ambiguity: Dynamic objects may remain static for several frames, and occlusions cause confusion between static background and dynamic foreground.
• Sensor artifacts: Sparse or noisy measurements hinder the reliable identification of motion.
• Computational efficiency: Real-time demand in robotics requires fast online solutions, often with limited access to semantic labels or global context.
• Trade-off: Overzealous removal may eliminate static structure (reducing preservation rate, PR), while under-removal leaves dynamic traces (lowering rejection rate, RR).
• Generalization: Domain, environment, and sensor variations require either highly adaptive methods or tunable pipelines.

These challenges motivate the diversity of approaches—ranging from handcrafted geometric tests, voxel-level occupancy schemes, region descriptors, to learned neural heuristics and statistical modeling.

2. Algorithmic Approaches

2.1 Geometric and Occupancy-Based Methods

Most classical algorithms for dynamic points removal are geometric or occupancy-driven.

• OctoMap uses occupancy grids with ray-casting to infer free space and thus remove points lying in known-free regions, but suffers from shallow-angle ray issues and computational expense (Zhang et al., 2023).
• Removert adopts visibility logic, tracing LiDAR returns through multi-resolution range images to eliminate points occluded by subsequent scans; it is fast but fails behind persistent occluders.
• ERASOR and ERASOR++ employ region-wise pseudo-occupancy, binning space in rings and sectors, measuring height differences or multi-bit height encodings, and using egocentric ratios or stack-based overlap tests to identify dynamic bins. This is followed by region-wise ground plane fitting to recover static ground and mark points above the plane as dynamic (Lim et al., 2021, Zhang et al., 8 Mar 2024). ERASOR++ incorporates binary height descriptors, Height Stack Test (HST), Ground Layer Test (GLT), and Surrounding Points Test (SPT) for improved performance.

2.2 Probabilistic and Bayesian Methods

• Height Interval Filtering (HIF) constructs a pillar-based representation, modeling the presence of static structure across dynamically-varying vertical intervals per grid column. Probabilities are updated via Bayesian filtering upon new scans (hit/miss events), and occlusions are mitigated by the Low-Height Preservation (LHP) strategy (2503.06863).
• FreeDOM leverages multi-resolution octrees with conservative free space estimation: only voxels observed as free both temporally and within a spatial neighborhood are cleared, avoiding false positives from transient occlusions. Dynamic segmentation thresholds are adaptive, and scan-integrated front-end is complemented by back-end map-refinement that clears residual dynamics using accumulated free space (Li et al., 15 Apr 2025).

2.3 Density-Based and Clustering Methods

• DynaHull applies convex hull volumetry in local k-neighborhoods, defining a per-point density factor. Dynamic points, likely to exhibit low density (large local convex hull for a given k), are eliminated by adaptive, cluster-wise thresholds after plane-based ground segmentation and k-means spatial partitioning (Habibiroudkenar et al., 15 Jan 2024).

2.4 Bitwise and Matrix Encoding Methods

• BeautyMap employs a binary-encoded adaptable ground matrix, where each vertical column of the spatial grid stores occupancy as a N-bit integer. Points present in the prior global map but missing in the scan trigger bitwise differencing; ground-aware hierarchical segmentation and a static-restoration module using visibility and reverse virtual ray-casting prevent over-removal of occluded static points (Jia et al., 12 May 2024).

2.5 Hash Maps and Real-Time Structures

• RH-Map uses a two-layer, region-wise hash map (region and cube layers) for online map management. The front-end uses region-wise ground plane estimation (R-GPE) and Scan-to-Map Removal (S2M-R) on each scan to suppress dynamic regions, while a lightweight back-end operates on keyframes to further prune residuals (Yan et al., 2023).

2.6 Learning-Based and Neural Methods

• PointCleanNet classifies outlier points by learning spatially localized characteristics from patches. Each point's local context is processed by a deep network, which outputs a removal score and displacement vector for further denoising. While originally designed for surface outliers, it is adaptable to dynamic removal with strong robustness to unseen noise (Rakotosaona et al., 2019).
• Dynamic Partial Removal (DPR) in LNS: In combinatorial optimization, dynamic points removal formalizes as a neural destroy operator (DPR) that probabilistically selects groups of nodes for removal using a Hierarchical Recurrent Graph Convolutional Network (HRGCN). This model learns adaptive, spatially-clustered, and temporally-consistent removal policies within a PPO reinforcement learning loop, directly optimizing solution exploration and repair (Chen et al., 2020).

3. Benchmarking, Evaluation, and Metrics

Comparative evaluation of dynamic points removal algorithms is organized around accuracy, computational efficiency, and robustness across datasets:

• Metrics: Standard measures include Static Accuracy (SA), Dynamic Accuracy (DA), Associated Accuracy (AA as geometric mean), Preservation Rate (PR), Rejection Rate (RR), and the harmonic mean F1-score (Zhang et al., 2023).
• Distance-error analysis: Histograms of the distance between false negatives and their nearest true positive dynamic point reveal the propensity of methods to miss dynamic points near ambiguity zones.
• Runtime: Efficiency ranges widely—from sub-10 ms per frame for schemes such as HIF (2503.06863) and label-consistency LIO (Yuan et al., 4 Jul 2024), to 50–1000 ms for voxel-based methods like OctoMap and variants (Zhang et al., 2023).
• Datasets: Common benchmarks are KITTI, SemanticKITTI, Argoverse 2.0, and domain-specific indoor datasets. These span semi-urban, dense city, and cluttered indoor sensor setups.
• Generalization: Benchmarks emphasize the importance of cross-sensor, cross-scenario robustness and the ability of algorithms to adapt their thresholds without manual tuning.

A summary of per-dataset benchmark results is shown below for several classical and recent algorithms (AA = Associated Accuracy):

Method	KITTI 00 (AA%)	KITTI 05 (AA%)	Argoverse 2.0 (AA%)	Semi-indoor (AA%)
Removert	64.3	47.1	55.5	34.9
ERASOR	81.1	82.9	87.7	79.3
OctoMap	82.4	81.1	79.8	85.5
OctoMap + G + F	95.8	93.0	82.6	84.3

4. Technical Innovations and Key Mechanisms

• Multi-resolution/Adaptive Grids: Methods such as FreeDOM and BeautyMap achieve high accuracy and efficiency by working with multi-scale or layered occupancy representations, facilitating both fine-grained removal and computational tractability (Li et al., 15 Apr 2025, Jia et al., 12 May 2024).
• Per-region or Per-pillar Reasoning: Algorithms like ERASOR, HIF, and BeautyMap exploit reg1onal or pillar-based encodings—height intervals, binary stacks, or height-coding descriptors—allowing for explicit modeling of spatial and vertical occupancy without global coordination (Lim et al., 2021, 2503.06863, Zhang et al., 8 Mar 2024).
• Probabilistic Filtering/Bayesian Updates: The incorporation of Bayesian updating in pillar interval probabilities, as in HIF, enables principled fusion of evidence from multiple scans, yielding robust filtering against false positives/negatives due to occlusion or noise (2503.06863).
• Neural Architectures: Leveraging GNNs to learn the context for adaptive removal (DPR/HRGCN) enables strong, context-aware heuristics for combinatorial search spaces, but comes with increased computational overhead per step and requires retraining for new distributions (Chen et al., 2020).
• Hybrid Pipelines: Split front-end/back-end designs (e.g., FreeDOM and RH-Map) support aggressive removal on-the-fly and correction of longer-term misclassifications post hoc with batch or keyframe processing, balancing precision with efficiency.

5. Limitations, Trade-offs, and Open Problems

• Ground-as-dynamic ambiguity: Algorithms that do not robustly segment or protect the ground often suffer from false dynamic classification in sloped, uneven, or occluded ground regions. Several papers recommend robust ground estimation as an essential initial phase (Zhang et al., 2023).
• Occlusion and Over-removal: Occlusion remains a persistent challenge, especially for visibility-based and occupancy flooding methods—dynamic artifacts can persist behind long-lived occluders or in regions lacking sufficient multi-view coverage (2503.06863, Jia et al., 12 May 2024).
• Parameter and Threshold Tuning: Methods relying on fixed (sensor-specific) height, occupancy, or removal thresholds may see significant performance drop if not retuned for new environments or sensor modalities. Adaptive or learned thresholds, as well as minimal-parameter pipelines, are highlighted as desirable (Zhang et al., 2023).
• Semantic and Learning-based Extension: While learning-based approaches (e.g., PointCleanNet, DPR) show promise, their generalization requires extensive retraining; hybrid strategies (semantic + geometric) may offer improvements but remain less explored in large-scale, real-time deployments.

6. Practical Applications and Performance

Dynamic points removal techniques are pivotal in:

• Robotics and SLAM: Real-time filtering of moving objects is necessary to prevent map corruption and ensure accurate localization, both in batch (mapping) and online (LiDAR-inertial odometry) settings (Yuan et al., 4 Jul 2024, Yan et al., 2023).
• High-fidelity Mapping: Urban and indoor mapping pipelines depend on dynamic-filtered point clouds to construct human-usable maps, support path planning, and enable long-term autonomy.
• Combinatorial Optimization: In vehicle routing or resource assignment, learned removal heuristics (DPR) allow scalable, adaptive search over large neighborhoods, with measurable improvement in solution quality and convergence (Chen et al., 2020).
• Downstream Perception: Denoised and cleaned representations improve downstream modules, e.g., object detection, tracking, or semantic segmentation, by avoiding confusion with moving artifacts.

Typical reported performance for state-of-the-art methods (PR = Preservation Rate, RR = Rejection Rate, F₁ = Harmonic mean):

Method	KITTI 00 F₁	KITTI 05 F₁	Indoor F₁	Mean Runtime
ERASOR	0.955	0.933	—	0.073 s / frame
ERASOR++	0.965	0.971	—	0.10–0.14 s
HIF	0.969	0.941	0.838	11.6 ms / frame
FreeDOM	0.996	0.983	0.990	16–90 ms / frame
BeautyMap	0.976	0.973	0.922	46 ms / frame
DynaHull	—	—	0.766 EMD	(offline-batch)

This table shows that recent approaches achieve >95% F1 on standard benchmarks, with real-time runtimes on commodity CPUs. HIF and FreeDOM notably yield 7–9x better runtime than prior arts with comparable or better accuracy (2503.06863, Li et al., 15 Apr 2025).

7. Prospective Directions and Recommendations

Recommendations drawn from comparative analyses and benchmarking studies include (Zhang et al., 2023):

• Start with robust ground plane segmentation, marking ground voxels in advance to insulate downstream dynamic-removal logic.
• Fuse multiple geometric and statistical cues (occupancy, visibility, pseudo-occupancy, surface intensity) to cross-validate potential dynamic/inconsistent regions.
• Incorporate adaptive, locally-refined thresholds by learning or calibrating from initial frames or from contextual distributions.
• Post-process with local neighborhood clustering to recover small or isolated clusters of missed dynamic points.
• Minimize the number and complexity of tunable parameters; design pipelines for easy transfer to new sensors/environments.
• Leverage hybrid architectures: lightweight, cross-frame map representations with scan-side dynamic pickup and global map correction in the back-end.
• Consider learning-based post-processing to refine the final dynamic/static labeling, especially in difficult or ambiguous regions.

Current research continues to pursue more generalizable, adaptive, and semantically-aware pipelines, as well as improved formal guarantees for convergence and real-time suitability in both indoor and outdoor, sparse and dense, and highly dynamic environments.