Image-Driven Multiview Environmental Simulation Modeling and Dynamic Response Mechanisms for Virtual Disaster Drills

1.png

Figure 1. Image-driven environmental simulation modeling process based on semantics-enhanced 3DGS

2.1 Point cloud space partitioning and appearance feature learning based on binary space partitioning (BSP) tree

Virtual disaster drills require extremely high geometric accuracy of the scene. For instance, subtle deformations in building load-bearing structures or the spatial layout of pipeline systems could directly impact disaster evolution simulation results. Although traditional 3DGS methods excel in rendering efficiency, the point cloud space they initialize lacks effective structural organization, leading to limited encoding ability for scene geometric details, especially in areas with texture loss or complex occlusion, where geometric distortion is common. To solve the core conflict between geometric reconstruction accuracy and physical property fidelity in large-scale complex scenes, this paper chooses to adopt BSP tree-based point cloud space partitioning in image-driven multi-view environmental simulation modeling for virtual disaster drills. As a spatial binary partitioning data structure, BSP trees can recursively divide unordered point clouds into spatial indices with a hierarchical organization. This structured representation provides a solid foundation for subsequent geometric optimization.

The BSP tree construction begins with systematically organizing the point cloud space generated by the Structure from Motion (SfM) algorithm. In practice, the system first merges the 3D point clouds generated from multiple images via SfM into a unified global spatial distribution map. Then, it samples this point cloud space by passing camera rays through it. Based on the computation of spatial axis variance, the system recursively divides the point cloud space into a left subtree and a right subtree until each leaf node contains a single point cloud. In the hierarchical structure of the BSP tree, this paper innovatively stores appearance attribute features at the nodes of the tree branches. Lower-level subtree nodes capture fine-grained local features, while higher-level nodes encode more macroscopic regional features. Each point cloud, based on its position index in the BSP tree, can quickly retrieve its corresponding color appearance feature D_Zu and geometric appearance feature D_hu.

To fully exploit the implicit features stored in the BSP tree, this paper designs a dedicated dual-branch Multi-Layer Perceptron (MLP) decoder architecture. The signed distance field (SDF) decoder accurately describes the geometric surface of the scene by estimating the SDF, with its ReLU activation function and 32 hidden nodes balancing computational efficiency and precision. Meanwhile, the RGB decoder is responsible for reconstructing the scene's color appearance, with 128 hidden nodes designed to capture complex texture details. These two decoders work together to convert the discrete point clouds into continuous scene representations via trilinear interpolation of point cloud features and positional encoding. Specifically, let the SDF decoder F_t be used to estimate $\hat{t}_o$, and the RGB decoder F_z be used to estimate the RGB color $\hat{z}_o$, where d is the position encoding function, as expressed in the following:

$\begin{gathered}\hat{t}_o=F_t\left(d_{(o)}, D_u(o)\right), \hat{z}_o=F_z\left(d_{(o)}, D_u(o)\right) \\ u=1,2, \ldots, M\end{gathered}$         （1）

During training, the system uses a multi-level supervision strategy to optimize the implicit representation of the BSP tree. The SDF branch uses binary cross-entropy loss (LOSS_BCE) combined with function regularization loss (LOSS_EIK) and smoothness loss (LOSS_SMOOTH) to ensure the continuity and accuracy of the geometric surface. The expressions are as follows:

$\begin{aligned} & \operatorname{LOSS}_{B C E}(o)=  T\left(t_o\right) \cdot \log \left(T\left(t_o\right)\right)+\left(1-T\left(t_o\right)\right) \cdot \log \left(1-T\left(t_o\right)\right)\end{aligned}$         （2）

$\operatorname{LOSS}_{E I K}(o)=\left(1-\left\|\nabla F_t\left(d(o), D_u(o)\right)\right\|\right)^2$        （3）

$\begin{aligned} & M_{\text {SMOOTH }}(o)=  \left\|\nabla F_t\left(d(o), D_u(o)\right)-\nabla F_t\left(d(o+\epsilon), D_u(o+\epsilon)\right)\right\|\end{aligned}$       （3）

The RGB branch uses the L1 loss function L₁ = |$\hat{z}_o$-z_o|, with the real colors from the input images as supervision signals. Notably, when determining the ground-truth color of the ground, the system selects the pixel color with the minimal Euclidean distance from the point cloud, effectively solving the color inconsistency issue between multi-view images.

Let η_EIK, η_SMOOTH, and η_RGB be the scaling factors. The global objective function for BSP tree implicit feature learning, LOSS(o), is defined as:

$\begin{aligned} & \operatorname{LOSS}_{(o)}=\operatorname{LOSS}_{\text {BCE }}(o)+\eta_{E I K} \operatorname{LOSS}_{E K}(o)  +\eta_{S M O O T H} \operatorname{LOSS}_{S M O O T H}(o)+\eta_{R G B} \operatorname{LOSS}_1(o)\end{aligned}$                          （5）

The advantages of applying this BSP tree-based point cloud space partitioning method in virtual disaster drill environment modeling are manifold. First, the hierarchical spatial index allows the system to flexibly adjust the reconstruction accuracy of different areas based on the specific drill requirements—finer divisions can be applied to critical facilities, while background areas can maintain a relatively sparse representation. Second, the rich appearance features stored in the BSP tree provide important input for subsequent physical simulations, such as the material reflection properties and surface roughness, which can be derived from these features. This structured scene representation provides an efficient query and update mechanism for dynamic disaster simulation. When a disaster causes scene changes, the system can quickly locate the affected areas and update the corresponding BSP tree nodes, enabling real-time evolution of the scene state.

2.2 Construction of 3D colored mesh with implicit feature fusion and dense supervision generation

Virtual disaster drills require extremely high geometric continuity of the scene. For example, in scenarios such as simulating stress distribution in building structures damaged by earthquakes or analyzing water flow paths during flood propagation, discrete Gaussian sphere representations are insufficient to provide continuous, complete surface information. To address the inherent limitations of traditional 3DGS methods in terms of geometric integrity and physical accuracy, this paper introduces a 3D colored mesh. The 3D colored mesh generated based on BSP tree implicit features can extract isosurface from the probability density function f(x) = 2e^-4 as a continuous mesh surface N₀ using the MarchingCubes algorithm:

$N_0=-2 \ln \left(\left(2 e^{-4}\right) \cdot(2 \pi)^{\frac{3}{2}}|\Sigma|^{\frac{1}{2}}\right)$              （6）

This explicit mesh representation provides an essential continuous geometric foundation for disaster physical simulation. Compared to sparse SFM point clouds, mesh representations can reconstruct geometric features that are lost due to sparse scanning, such as complete walls, pipeline systems, and other key structures.

At the technical implementation level, the construction of the 3D colored mesh is based on the BSP tree implicit features learned in the previous steps. Using the SDF values and RGB colors predicted by the dual-branch MLP decoder, the system uses the MarchingCubes algorithm to perform isosurface extraction in the 3D space composed of BSP tree leaf nodes. The algorithm traverses each small cube, determining its intersection with the isosurface based on the SDF values at the cube's vertices. It then computes the position and color attributes of the intersection points through linear interpolation, ultimately connecting these points into triangular faces, forming a complete mesh model that includes both geometric information and color textures. Specifically, let the original point cloud space R be partitioned into R' with multiple layers of leaf nodes by the BSP algorithm, the camera motion path provided by the original image shooting path be S, and the input images be U, then a set of density depth images can be generated from the constructed 3D mesh L:

$F=\operatorname{RayTracing}\left(L \mid S, R, U, R^{\prime}\right)$            （7）

This process converts discrete point cloud data into continuous surface representations while maintaining the scene's visual fidelity.

The core advantage of the 3D-colored mesh lies in its provision of dense geometric and appearance supervision for 3DGS optimization. Unlike traditional 3DGS, which only uses sparse SFM point clouds for deep supervision, this method generates a set of dense depth images from the rendered 3D mesh using a ray-tracing algorithm. These depth images, consistent with the resolution of the input images, provide richer and continuous geometric constraints. During training, the dense depth supervision effectively prevents the Gaussian spheres from excessive expansion or contraction in textureless areas, significantly improving the accuracy of the reconstructed geometry. At the same time, the RGB color information stored at the mesh vertices provides more reliable guidance for the appearance optimization of Gaussian spheres, ensuring consistent material texture representation.

2.3 Semantics-enhanced GS and model optimization based on 3D mesh priors

Virtual disaster drills not only require visual realism in scenes but also need to ensure geometric accuracy in representation and feasibility in dynamic updates. To solve the fundamental limitations of traditional 3DGS methods in geometric consistency and physical property representation, this paper employs GS based on implicit features and 3D colored mesh as the final optimization step. Traditional 3DGS directly initializes Gaussian spheres from sparse SFM point clouds, which tends to produce geometric distortions in textureless areas or complex structures. In contrast, this method constrains the initialization of Gaussian spheres in a physically plausible spatial distribution using the continuous surface prior provided by 3D mesh vertices. This mesh-based initialization strategy is especially suitable for precise reconstruction of key structures in disaster scenarios, providing a solid geometric foundation for subsequent physical simulations.

At the technical implementation level, this method innovatively performs deep fusion of BSP tree-learned implicit features with explicit geometry from 3D meshes. The 3D mesh vertices not only provide positional information ω but also serve as carriers of rich attribute information. By retrieving implicit features D(n) from the BSP tree, geometric appearance and color appearance priors are injected into each Gaussian sphere. The introduction of a feature encoder further enhances the expressiveness of this process by encoding the average position ω, scale t, rotation w, color z, and the newly added attributes D_u(n) into a unified feature representation, integrated into the spherical harmonic coefficients tg and opacity p of the Gaussian spheres. Specifically, the properties of the 3D Gaussian spheres are initialized as:

$\Gamma=\left\{\begin{array}{c}\omega(n), t(n), w(n), z(n) \\ \mid L ; p\left(D_u(n), \operatorname{tg}\left(z(n), D_u(n)\right), D_u(n) \mid L, D\right)\end{array}\right\}$           （8）

$u=1, \ldots, M$

The mesh-based attribute initialization fundamentally enhances the physical authenticity of GS. In virtual disaster drills, physical properties such as material thermal conductivity, combustibility, and structural strength directly affect disaster evolution processes. By associating the implicit features at the mesh vertices with the spherical harmonic coefficients of the Gaussian spheres, the system is able to encode the microscopic optical properties of materials, such as high-gloss reflection for metals and diffuse reflection properties for wood. These optical properties are intrinsically linked to physical parameters, for example, surface roughness affecting flame adhesion capacity, and color saturation reflecting material moisture content. Therefore, the optimized set of Gaussian spheres not only ensures visual realism but also forms a computable physical field, providing rich input parameters for disaster simulation.

In the rendering and optimization phase, the method uses differentiable rasterization and adaptive density control mechanisms to ensure stability during the training process and realism in the final result. By indexing Gaussian spheres in depth order within camera space, the system achieves efficient foreground-background blending, where the covariance matrix in the β blending formula accurately describes the anisotropic spatial distribution. The specific covariance matrix expression is:

$\sum^{\prime}=K R \sum R^S K^S$        （9）

The color Z, depth F, and total opacity P for each pixel can be computed as:

$\begin{aligned} & Z=\sum_{u=1}^V S_u \beta_u z_u, F=\sum_{u=1}^V S_u \beta_u z_u \\ & P=\sum_{u=1}^V S_u \beta_u z_u, S_u=\prod_{k=1}^{u-1}\left(1-\beta_k\right)\end{aligned}$            （10）

β is computed as: β_u = p_u·exp(-1/2(a- ω_u)^SΣ'^-1(a-ω_u)), and for a specific view (R, U) in the scene containing V 3D Gaussian spheres, the image U and depth F are rendered via the differentiable rendering function E:

$\hat{U}, \hat{F}=E\left(\left\{\Gamma_u, u=1,2, \ldots, V\right\} \mid R, U\right)$            （11）

Here, the use of inverse depth enhances numerical stability, making it especially suitable for rendering large-scale outdoor scenes. During training, adaptive density control dynamically prunes or clones Gaussian spheres based on gradient information in the view space. This feature is particularly important for dynamic updates in disaster scenarios—when building structures are damaged, the system can quickly adjust the distribution of Gaussian spheres in the affected areas, reflecting real-time changes in the scene. Assuming the scaling factors η_SS, η_DEP, and η_SMOOTH are defined, the original RGB image is denoted by U, the density depth rendered by the 3D mesh is denoted by F, and the structural similarity index between the original RGB and rendered RGB images is represented by LOSS_SS. The inverse depth smoothness term between the original RGB image and rendered depth is represented by LOSS_SMOOTH. The following equation gives the objective function used during training:

$\begin{aligned} & \operatorname{LOSS}=\left(1-\eta_{S S}\right) \operatorname{LOSS}_1(U, \hat{U})+\eta_{D E P} \operatorname{LOSS}_{S S}(F, \hat{F}) +\eta_{S M O O T H} \operatorname{LOSS}_{S M O O T H}(U, \hat{F})\end{aligned}$                             （12）

The innovative application of these methods in virtual disaster drills shows value in multiple dimensions. Geometrically, the Gaussian sphere distribution initialized based on the mesh ensures the structural integrity and continuity of the scene, making physical calculations such as collision detection and ray tracing more accurate and reliable. In terms of materials, the enhanced Gaussian spheres with implicit features can distinguish the behavior characteristics of different materials during disasters, such as the different reaction modes of concrete and glass under heat. In terms of dynamic response, the structured Gaussian sphere representation supports a local update mechanism, allowing the system to quickly reconstruct the Gaussian sphere parameters of affected areas when disasters such as explosions or collapses occur, achieving real-time evolution of the scene state.

4.png

Figure 4. Dynamic efficiency evaluation of personnel evacuation in disaster environments based on AI agent models

To verify the effectiveness of the dynamic response mechanism in the virtual drill environment built in this paper, we conducted a simulation experiment on personnel evacuation in a disaster environment using an AI agent model, to evaluate the real-time efficiency of different evacuation paths. The experimental results are shown in Figure 4. The four evacuation flows exhibit significant dynamic characteristics: the main exit evacuation flow in the east zone quickly peaks at approximately 20 persons/min within the first 50 minutes but sharply drops to below 11 persons/min after 100 minutes, indicating that while the early capacity of this exit is strong, it is prone to bottlenecks; the emergency passage evacuation flow in the west zone shows a steady upward trend, gradually increasing from 10 persons/min to 23 persons/min, demonstrating its sustained reliability as a secondary route; the underground to upper evacuation flow fluctuates throughout the process, reflecting the complexity of vertical evacuation affected by building structure; while the cross-regional transfer flow remains inactive throughout the entire simulation period, reflecting the simulation model's decision-making ability to dynamically adjust paths according to environmental changes.

To systematically evaluate the comprehensive performance of different modeling methods in virtual disaster drill scenarios, we designed a multidimensional evaluation system, including geometric accuracy, visual fidelity, semantic integrity, and reconstruction completeness, and conducted comparative experiments on a unified dataset. As shown in Figure 5, the proposed semantic-enhanced 3DGS method performs excellently in the overall modeling quality score: during the initial training phase, our method already achieves a comprehensive score of nearly 34, significantly higher than the 31 points for Point-Based Rendering and 29 points for Neural Feature Fields. As training progresses, our method reaches a stable peak score of 37 points after 15,000 iterations, while 3D-GS requires 25,000 iterations to reach 36 points, and the point-based rendering method ultimately only achieves 32 points. This result proves that through the introduction of BSP tree implicit features and 3D mesh priors, our method achieves significant improvements in all dimensions of modeling quality while maintaining training efficiency. These experimental results fully validate the practical value of the research in virtual disaster drills. High-quality environment modeling is the foundation for the reliable operation of subsequent dynamic response mechanisms—a comprehensive score of 37 indicates that the reconstructed scene is not only visually realistic but also maintains a complete geometric structure and semantic information, which is crucial for the accuracy of disaster simulations.

To more precisely evaluate the model's performance for this specific task, relevant experiments were set up in this paper. The data from Table 3 systematically verifies the significant advantages of the proposed method in the environmental simulation modeling for virtual disaster drills, from the perspectives of visual fidelity, geometric accuracy, and structural semantic integrity. In traditional image quality metrics, the proposed method leads comprehensively: with the highest values of PSNR 28.53 dB and SSIM 0.868, it indicates the optimal pixel-level color and structural similarity of the reconstructed scene. The lowest LPIPS of 0.231 demonstrates that the rendered results have the smallest perceptual difference from real images, effectively avoiding unnatural artifacts and distortions. This advantage ensures a high degree of visual immersion in the virtual drill environment, which is crucial for maintaining the participants' sense of presence and focus. However, for disaster drills, visual quality is just the foundation, and deeper requirements are geometric and physical accuracy.

In terms of geometric accuracy and structural integrity, the proposed method provides a reliable digital foundation for physical simulations, which is core to supporting the "dynamic response mechanism." The proposed method significantly outperforms all comparison models in geometric accuracy, being 3.2 percentage points higher than the next best model, 3D-GS. This proves that by introducing BSP tree and 3D mesh priors, the proposed method can more accurately restore the 3D geometric form of building structures, such as the precise dimensions of load-bearing columns, the slope and steps of stairs, and the flatness of walls. This millimeter-level geometric fidelity is an absolute prerequisite for subsequent precise physical calculations. The structural integrity score of 86.4% is the key metric that most reflects the value of the proposed method. This score means the reconstructed model not only accurately reflects the visible surface but can also infer and complete the structural logic in accordance with the physical laws of the real world, such as wall continuity and the connection relationships between beams and columns. A model that scores high in "structural integrity" ensures that force transfer and structural damage in dynamic disaster simulations follow physical laws, rather than producing illogical "floating" or "fracturing," greatly improving the scientific and credible nature of disaster simulation results.

5.png

Figure 5. Performance comparison of different modeling methods on the disaster scenario dataset

Table 6. Indoor fire risk assessment experimental results