Object-Free Environment Map Estimation

• Object-free environment map estimation is defined as techniques that reconstruct static maps by excluding transient objects to accurately capture free space, walls, and lighting.
• Approaches employ probabilistic occupancy grids, Gaussian models, and dynamic removal methods to enhance segmentation and map fidelity in complex scenes.
• Integration with SLAM and VIO frameworks improves computational efficiency and localization accuracy by decoupling static structure estimation from dynamic interference.

Object-free environment map estimation refers to techniques designed to construct maps of an environment that are minimally contaminated by transient, dynamic, or moveable objects. These methods target the extraction or inference of static structural properties—such as free space, walls, or global environmental lighting—without explicit modeling, storage, or influence from non-structural content (e.g., vehicles, furniture, or people). Applications range from autonomous navigation, SLAM, and room segmentation, to mixed reality and lighting estimation. This article surveys core algorithms, representative methodologies, practical toolchains, and typical challenges and solutions, as reflected in recent research.

1. Probabilistic and Grid-Based Free Space Estimation

One foundational family of object-free mapping methods is stochastic occupancy grid estimation, exemplified by probabilistic tessellation of the environment into discrete cells encoding likelihoods of occupancy. Using stereoscopic depth extraction, each grid cell accumulates occupancy evidence from 3D measurements projected onto a ground plane. Gaussian models are applied to quantify measurement uncertainty: $$G_{mx}(S_k) = \frac{1}{(2\pi)^{1/2} |T_k|^{1/2}} \exp\left( -\frac{1}{2} (\varepsilon_k^T T_k^{-1} \varepsilon_k) \right)$$ where $T_k$ is the measurement covariance and $\varepsilon_k$ is the error vector. For each grid cell $(i,j)$, the occupancy likelihood $L_{ij}(m_k)$ of measurement $m_k = (u_k, v_k, d_k)$ is calculated: $$L_{ij}(m_k) = G_{mx}((u_{ij} - u_k,\, 0,\, d_{ij} - d_k))$$ Aggregating over all $m$ measurements, the occupancy likelihood is

$D(i, j) = \sum_k L_{ij}(m_k)$

Segmentation then identifies the free space with column-wise thresholding, marking the first cell as the boundary and occluding those behind it. This approach robustly accounts for uncertainty and noise in measurement, as demonstrated on urban driving scenarios using the KITTI dataset (Sahdev, 2017).

2. Dynamic Object Removal and Map Purification

In realistic environments, dynamic objects introduce artifacts (e.g., ghost vehicles, transient pedestrians) that must be excluded from static maps. Solutions bifurcate into real-time scan-based removal and offline map refinement:

• Scan-based removal: Multi-resolution map structures facilitate rapid segmentation of LiDAR scans by estimating free space at coarse voxel levels and static space at subvoxel levels. Raycast enhancement supplements missing background observations (e.g., sky), improving segmentation when raw returns are absent. Each scan point is assigned a DynamicLevel based on proximity to free space and neighbor support.
• Map refinement: Residual dynamic artifacts are addressed by incrementally updating occupancy evidence, for example using counters:

$n_f^{(t)} = n_f^{(t-1)} + 1, \qquad n_o^{(t)} = 0$

Voxel is marked as free if all spatial neighbors meet a minimum free-count threshold. Subvoxels track timestamps and dynamic levels, enabling time-based and spatial neighborhood clearing. These two-stage systems achieve an average F1-score improvement of 9.7% over previous visibility-based methods (Li et al., 15 Apr 2025).

• Pseudo occupancy concept: ERASOR (Lim et al., 2021) defines pseudo occupancy as a bin’s vertical extent, allowing scan ratio tests to flag candidate bins for dynamic contamination:

$$\Delta h_{i,j,t} = \sup \{ Z_{i,j,t} \} - \inf \{ Z_{i,j,t} \}$$

Followed by region-wise ground plane fitting and seed-based plane growing using PCA, dynamic points are efficiently removed without the ambiguity and overhead of classical raytracing or visibility approaches.

3. Semantic and Structure-Free Mapping

Object-free methods extend to environments with high clutter (indoors), where static structure (walls, ceilings, doors) must be preserved at the expense of transient or moveable entities. A SLAM-based dual-LiDAR system uses a vertical scanner for semantic segmentation and wall plane detection, aided by context-sensitive rules for doors and a robust RANSAC plane growing algorithm (He et al., 2019). The process consists of:

• Point cloud rearrangement: Filtering out floor and ceiling, then ordering points by laser beam angle.
• Wall plane detection: Computing forward difference on vertical lines, thresholding to find planes, iteratively merging via parameter similarity.
• Semantic labeling: Assigning major categories, including automated door detection by local wall recess analysis.

Reported wall precision of 99.60% confirms reliable extraction of immobile environmental structure, supporting downstream tasks in room segmentation and long-term localization.

4. Object-Free Environment Map Estimation for Navigation and SLAM

Advanced SLAM and visual-inertial odometry (VIO) frameworks increasingly leverage object-free formulations to improve computational efficiency and robustness. The structureless VIO approach removes explicit map points (3D features/inverse depths) from the state vector, instead relying solely on the epipolar constraint between bearing vectors across keyframes: $$r_{ij}^n(x_{c_i}, x_{c_j}) = \left( {}_{I_j}^G R_C^I z_j^n \right)^T [ t / \|t\| ]_\times \left( {}_{I_i}^G R_C^I z_i^n \right)$$ Here $t$ is the translation vector between camera frames. This eliminates depth estimation, reducing optimization complexity and decoupling localization from structure recovery. On benchmark datasets, average solve time decreased (EuRoC: 35.01 ms $\to$ 14.97 ms) and trajectory error improved (ATE: 0.195 m $\to$ 0.168 m) compared to structure-based VIO (Song et al., 18 May 2025). Such map-free localization approaches offer robust performance in feature-sparse or corridor scenes and simplify data association.

5. Lighting and Radio Environment Map Estimation

Map estimation without explicit object modeling is also essential in global environmental understanding (lighting, radio signal strength).

• Lighting map inference: Deep learning methods recover HDR illumination maps from LDR spherical panoramas by predicting spherical harmonic (SH) coefficients, bypassing the need for exhaustive scene geometry or ground truth lighting data. A two-stage network—an LDR-to-HDR autoencoder and a lighting encoder—regresses SH coefficients per color channel, regularized by a spectral prior:

$$E(\rho) \approx \sum_{l=0}^2 \sum_{m=-l}^l L_m^l Y_m^l(\rho)$$

Supervision employs a differentiable relighting operator and photometric loss, grounded in a global Lambertian assumption. Imposing a structured prior improves accuracy by ~50% (m-RMSE: 0.0101 vs 0.0229). Applications span mixed reality, virtual reality, and VFX compositing (Gkitsas et al., 2020).

• Radio environment mapping: Graph Neural Networks are employed to estimate spatially continuous radio signal maps (RSRP/RSRQ) without explicit modeling of obstacles or objects. Cities are partitioned into H3-indexed hexagonal tiles, features are aggregated via GCN layers, and signal metrics are predicted per tile. The method achieves R² scores of 0.83 and classification accuracy of 92%, outperforming tabular and fully connected baselines (Shibli et al., 9 Jun 2024).

6. Comparative Analysis, Challenges, and Applications

Object-free map estimation methods must balance efficiency, fidelity, and robustness to dynamic interference and measurement uncertainty. Comparative findings highlight:

Method	Dynamic Object Handling	Static Structure Precision	Computational Efficiency
FreeDOM (Li et al., 15 Apr 2025)	Scan-based + map refinement	F1-score $\uparrow$ 9.7%	Real-time, multi-sensor support
ERASOR (Lim et al., 2021)	Pseudo occupancy + R-GPF	PR $>$88–94%, RR $>$95%	$<$0.07s per iteration
Structureless VIO (Song et al., 18 May 2025)	No explicit map, epipolar residuals	ATE improvement (0.195m $\to$ 0.168m)	Halved solve time (EuRoC)
Furniture-Free (He et al., 2019)	Vertical scan + semantic labeling	99.60% wall precision	Real-time (tens of ms per scan)

A plausible implication is that object-free frameworks provide significant performance advantages in domains requiring reliable static structure extraction, long-term localization, topological mapping, mixed-reality contextualization, and autonomous robot navigation. However, limitations may arise from sensor noise, SLAM drift, or ambiguous dynamic signatures. Continued research in incremental refinement strategies, semantic segmentation, and spectral regularization is likely to enhance robustness and generalization.

7. Future Implications and Directions

Recent trends indicate increasing deployment of multi-resolution, visibility-free, and deep learning-based approaches for object-free map estimation. Emphasis on computational efficiency, sensor-agnostic processing (LiDAR, stereo, radio, VR/AR cameras), and real-time operation underpins the adoption in autonomous vehicles, mobile robotics, and large-scale semantic mapping. Ongoing work aims to improve handling of edge cases (e.g., occlusion by persistent objects, drift in pose estimation), integration of diverse sensor streams, and adaptation to complex environments (urban, indoor, multimodal). Developing standardized benchmarks and performance metrics remains crucial for comparative evaluation and continued advancement in the field.