A virtual reality platform for 3D scene reconstruction and investigation simulation in criminal cases

Abstract

This invention belongs to the field of criminal investigation 3D scene reconstruction and judicial deduction technology. It discloses a virtual reality platform for criminal investigation 3D scene reconstruction and investigation deduction, comprising: fusing cross-modal evidence from the scene, anchoring spatiotemporal coordinates using a BeiDou-blockchain dual-reference system, detecting evidence conflicts, constructing an adaptive arbitration factor based on a forensic parameter library, and generating an evidence fusion package; constructing a resource slicing strategy to divide four-dimensional resource levels, converting text descriptions into skeletal motion topology chains, constructing a simplified spatiotemporal trajectory set, synchronously triggering case deduction, and generating a resource arbitration log; constructing a physical deduction rule set, then executing rainstorm erosion correction to generate a physical deduction instruction set; performing case deduction resource scheduling, generating a 3D deduction animation stream, constructing a judicial causal chain, and generating a crime logic analysis report; hot-loading desensitization rules to identify and replace sensitive fields; quantifying the effectiveness of the deduction, and optimizing and updating the forensic parameter library and resource slicing strategy.

Description

Technical Field

[0001] This invention relates to the field of criminal investigation three-dimensional scene reconstruction and judicial deduction technology, and more specifically, this invention relates to a virtual reality platform for criminal investigation case three-dimensional scene reconstruction and investigation deduction. Background Technology

[0002] In the field of 3D reconstruction of criminal cases, although existing technologies can achieve basic scene modeling, some problems still exist, such as cross-modal data fragmentation, coarse-grained spatiotemporal extrapolation, and lack of physical rules. Multi-source evidence (point cloud / text / video) lacks a unified spatiotemporal benchmark and arbitration mechanism, causing the reconstruction model to deviate from the real scene. Case extrapolation is limited by time accuracy and cannot restore key behavioral chains. Furthermore, due to the lack of integration of forensic biomechanical rules and environmental dynamic response, physical simulations such as bloodstain diffusion and trauma formation are seriously distorted, directly affecting the legal validity of investigation conclusions.

[0003] Current mainstream solutions have some drawbacks: cross-modal fusion methods based on manual calibration are inefficient and highly subjective, and cannot automatically arbitrate conflicts of physical evidence; spatiotemporal modeling often uses fixed time slices, making it difficult to capture the millisecond-level causal chain of the cause of the incident; the physics engine relies on game-level simplified algorithms and does not introduce a forensic parameter library and environmental correction mechanism, resulting in excessive errors in the deduction results under environmental influences, and the resource scheduling, behavior synthesis, and judicial verification links are completely separated, forming a broken chain of "reconstruction-deduction-verification", which hinders the closed-loop optimization of investigation decisions. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a virtual reality platform for three-dimensional scene reconstruction and investigation simulation in criminal cases, comprising:

[0005] Cross-modal evidence fusion unit: It fuses cross-modal evidence from the scene, uses BeiDou-blockchain dual-reference anchor spatiotemporal coordinates to detect conflicts in physical evidence, and generates adaptive arbitration factors based on a forensic parameter library to construct a physical evidence fusion package;

[0006] Resource Slicing and Trajectory Compression Unit: Parses the physical evidence fusion package, constructs a resource slicing strategy to classify four-dimensional resources; transforms text descriptions into skeletal motion topology chains, constructs a simplified set of spatiotemporal trajectories; synchronously triggers case deduction, and generates resource arbitration logs;

[0007] Forensic physical rule injection unit: Based on the simplified set of spatiotemporal trajectories and resource arbitration logs, a set of physical deduction rules is constructed; then, rainstorm erosion correction is performed to generate a set of physical deduction instructions;

[0008] Criminal investigation behavior synthesis and tracing unit: Based on the physical deduction instruction set, it performs case deduction resource scheduling and generates a three-dimensional deduction animation stream; then it constructs a judicial causal chain, binds it to a spatiotemporal version, and generates a crime logic analysis report;

[0009] Judicial Credibility Iterative Unit: Based on the 3D simulation animation stream and crime logic analysis report, hot-load desensitization rules are applied to identify and replace sensitive fields; the effectiveness of the simulation is quantified, and the forensic parameter library and resource slicing strategy are optimized and updated.

[0010] Furthermore, the method of detecting evidence conflicts by anchoring spatiotemporal coordinates using the BeiDou-Blockchain dual-reference point includes:

[0011] Integrate on-site laser point clouds, text reports, and surveillance video streams as cross-modal evidence;

[0012] Map the local coordinates of the laser point cloud to the BeiDou positioning system to establish a global three-dimensional spatial benchmark; align the timestamps of the monitoring video stream with the blockchain timestamps.

[0013] The relative position description in the text report is loaded from the pre-set forensic parameter library and converted into three-dimensional coordinates. These coordinates are then fused with the calibrated laser point cloud and surveillance video stream to generate a spatiotemporally aligned original evidence package.

[0014] Furthermore, the generation method of the evidence fusion package includes:

[0015] Based on the original evidence package, semantic conflicts are identified by comparing the coordinates of the same timestamp in the surveillance video stream and the laser point cloud.

[0016] Extract numerical fields from the text report and compare them with the geometric measurement results of the laser point cloud to identify numerical conflicts;

[0017] For numerical conflicts, an adaptive arbitration factor is constructed based on the forensic parameter library. If the numerical conflict meets the expectations, the adaptive arbitration factor is accepted for correction; if the numerical conflict does not meet the expectations, the dynamic correction process is initiated.

[0018] For semantic conflicts, the arbitration rule is defined as prioritizing the use of standard values ​​from the forensic parameter library, and using laser point cloud coordinates when standard values ​​are missing;

[0019] By combining conflict resolution records and reliability allocation results, an evidence fusion package is generated.

[0020] Furthermore, the method for constructing the resource slicing strategy to divide the four-dimensional resource levels includes:

[0021] Based on the physical evidence fusion package, a resource slicing strategy is constructed to quantify spatiotemporal density, physical evidence value, behavioral complexity, and environmental interference as four-dimensional resource characteristics, and these are further divided into different levels to obtain four-dimensional resource levels.

[0022] The four-dimensional resource level is mapped to a numerical value, and the GPU resource preemption priority is obtained by combining the four-dimensional resource characteristics.

[0023] GPU resources are dynamically allocated based on GPU resource preemption priority, and a resource slicing strategy table is generated.

[0024] Furthermore, the methods for constructing a simplified set of spatiotemporal trajectories and synchronously triggering case simulations to generate resource arbitration logs include:

[0025] Based on the evidence fusion package and the optimized resource slicing strategy table, the language description in the text report is converted into an action chain, and the action attributes are labeled to generate a skeletal action topology chain.

[0026] The three-dimensional space of the laser point cloud is voxelized and mapped to the voxel index using BeiDou coordinates. Then, through time slicing, a simplified set of spatiotemporal trajectories is constructed by combining the skeletal motion topology chain.

[0027] Based on the GPU resource preemption priority, the case is simulated, and GPU resource preemption events during the case simulation are recorded synchronously to generate a resource arbitration log.

[0028] Furthermore, the construction method of the physical deduction rule set includes:

[0029] Based on the simplified set of spatiotemporal trajectories, blood rheological parameters are loaded, and the corresponding trauma parameter table is obtained by querying the forensic parameter library.

[0030] Then, the trauma parameter table is mapped to physical boundary conditions, and blood rheological parameters are fused to construct a set of physical deduction rules.

[0031] Furthermore, the generation method of the physical deduction instruction set includes:

[0032] Match environmental records with the timestamp corresponding to the case from the external environmental parameter database to obtain the risk of impact from rainstorm erosion;

[0033] Based on the set of physical deduction rules and combined with the risk of impact from heavy rain erosion, the area of ​​bloodstain diffusion was corrected.

[0034] Simultaneously, rainstorm levels are classified based on the risk of erosion from heavy rain, and the wound depth of standard wounds is dynamically adjusted according to the rainstorm level; then, a set of physical deduction instructions is generated.

[0035] Furthermore, the generation method of the 3D simulation animation stream includes:

[0036] Based on the GPU utilization rate in the resource arbitration log and the GPU resource preemption priority of each case, the computation scheduling strategy for case simulation is dynamically adjusted.

[0037] Based on the adjusted computation scheduling strategy, the skeletal motion topology chain is completed and used as motion instructions. Combined with the physical deduction instruction set, a frame-level behavior description is generated, thus obtaining a 3D behavior deduction animation stream.

[0038] Synchronously record the GPU usage status and computing scheduling strategy for the current case, and update the resource arbitration log.

[0039] Furthermore, the method for generating the crime logic analysis report includes:

[0040] Based on the 3D behavior deduction animation flow, combined with the physical deduction instruction set and the spatiotemporal trajectory simplification set, physical evidence is associated with action behavior to construct a causal logic chain of criminal behavior, and then the causal chain conclusion is obtained.

[0041] Based on the causal chain conclusion, four-dimensional spatiotemporal coordinates are bound to the causal logic chain to generate spatiotemporal binding tags, which are then integrated to generate a crime logic analysis report.

[0042] Furthermore, the methods for optimizing and updating the forensic parameter library and resource slice strategy include:

[0043] Identify sensitive fields in the 3D simulation animation stream and crime logic analysis report, and hot-load the desensitization rules to desensitize and replace the sensitive fields;

[0044] A judicial credibility mechanism is established to quantify the degree of matching between the deduction results and the actual evidence in the physical evidence fusion package, thereby obtaining the credibility of the case deduction results;

[0045] Based on the reliability results, the forensic parameter library is updated locally and synchronized to the central server for global aggregation, resulting in an updated forensic parameter library and an optimized resource slicing strategy.

[0046] The technical effects and advantages of the virtual reality platform for 3D scene reconstruction and investigation simulation in criminal cases of this invention are as follows:

[0047] This invention systematically addresses some shortcomings in the background technology through a five-step closed-loop architecture driven by judicial cognition:

[0048] First, an immutable spatiotemporal coordinate system is constructed by anchoring the BeiDou-Blockchain dual benchmarks. Then, by combining conflict arbitration driven by the forensic parameter database with federally weighted confidence fusion, millimeter-level alignment of multi-source evidence is achieved, eliminating spatial distortion and semantic ambiguity in the reconstruction model from the source.

[0049] Secondly, the system classifies resources into four levels: spatiotemporal density, physical evidence value, behavioral complexity, and environmental interference, and generates a millisecond-level concise set of spatiotemporal trajectories (a four-dimensional spatiotemporal cube). Through a dynamic GPU preemption strategy, the system ensures zero-latency response of computing resources for high-priority cases such as bombings, thus breaking through the bottleneck of deduction efficiency.

[0050] Then, the bloodstain rheology Navier-Stokes model and trauma dynamics parameter library are loaded, and environmental adaptive correction (such as rainstorm erosion coefficient compensation) is performed to ensure that the physical simulation strictly conforms to forensic standards and reduces the inference error in complex environments to a range that can be legally accepted.

[0051] Next, based on the RPM behavior synthesis engine, the suspect's continuous actions are generated, a judicial causal chain of "physical evidence-behavior-motive" is constructed and bound with spatiotemporal version tags, and dual-channel evidence of 3D animation and logical report is output, realizing the visualization of investigation deduction and the unity of legal logic;

[0052] Finally, privacy compliance is ensured through a zero-knowledge proof desensitization mechanism, and a three-dimensional weighted evaluation system of physical evidence consistency, behavioral rationality, and environmental responsiveness is constructed to quantify the credibility of the deduction. Based on federated learning, the forensic parameter library and resource strategy are optimized in reverse, forming a complete autonomous evolution system for criminal investigation deduction. This changes the traditional investigation model of "human experience-driven and fragmented technical links" and provides a full-chain technical foundation for smart justice. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of a virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to the present invention;

[0054] Figure 2 This is a flowchart illustrating the resource slicing and trajectory compression unit in a virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to the present invention.

[0055] Figure 3 This is a schematic diagram of a virtual reality method for three-dimensional scene reconstruction and investigation simulation of criminal cases according to the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1

[0058] Please see Figure 1 and Figure 2 As shown in the figure, this embodiment provides a virtual reality platform for three-dimensional scene reconstruction and investigation simulation in criminal cases, including:

[0059] Cross-modal evidence fusion unit: It fuses cross-modal evidence from the scene, uses BeiDou-blockchain dual-reference anchor spatiotemporal coordinates to detect conflicts in physical evidence, and generates adaptive arbitration factors based on a forensic parameter library to construct a physical evidence fusion package;

[0060] Resource Slicing and Trajectory Compression Unit: Parses the physical evidence fusion package, constructs a resource slicing strategy to classify four-dimensional resources; transforms text descriptions into skeletal motion topology chains, constructs a simplified set of spatiotemporal trajectories; synchronously triggers case deduction, and generates resource arbitration logs;

[0061] Forensic physical rule injection unit: Based on the simplified set of spatiotemporal trajectories and resource arbitration logs, a set of physical deduction rules is constructed; then, rainstorm erosion correction is performed to generate a set of physical deduction instructions;

[0062] Criminal investigation behavior synthesis and tracing unit: Based on the physical deduction instruction set, it performs case deduction resource scheduling and generates a three-dimensional deduction animation stream; then it constructs a judicial causal chain, binds it to a spatiotemporal version, and generates a crime logic analysis report;

[0063] Judicial Credibility Iterative Unit: Based on the 3D simulation animation stream and crime logic analysis report, hot-load desensitization rules are applied to identify and replace sensitive fields; the effectiveness of the simulation is quantified, and the forensic parameter library and resource slicing strategy are optimized and updated.

[0064] Methods for detecting conflicts in physical evidence by anchoring spatiotemporal coordinates using the BeiDou-Blockchain dual-reference system include:

[0065] Three-dimensional spatial data of the crime scene is collected by a 3D laser scanner and used as a laser point cloud, which includes the local spatial coordinates (x, y, z) of the crime scene, the normal direction of the object surface, and the reflection intensity.

[0066] The forensic descriptions of the wounds and physical evidence on the body (such as "the wound is located on the left shoulder and is 3 cm deep"), as well as the descriptions of the scene investigation by professionals, are compiled into a text report containing natural language descriptions.

[0067] The real-time video captured by surveillance cameras around the scene is used as a surveillance video stream, including timestamps and video frame content (such as the suspect's movement trajectory, the time when physical evidence appeared, etc.).

[0068] Integrate on-site laser point clouds, text reports, and surveillance video streams as cross-modal evidence;

[0069] The BeiDou API is used to obtain the geographical coordinates (latitude, longitude, and altitude) of the crime scene. The local coordinates of the laser point cloud are mapped to the BeiDou positioning system through a coordinate transformation algorithm to establish a global three-dimensional spatial benchmark. The Aisino blockchain node is called to obtain millisecond-level timestamps and the timestamps of the surveillance video stream are time-aligned with the blockchain timestamps.

[0070] One method for time alignment is to take the difference between the blockchain timestamp and the timestamp of the surveillance video stream. If the difference does not meet expectations (e.g., >5ms), dynamic compensation is triggered by the PTP protocol.

[0071] The standardization process for laser point clouds involves reading laser point cloud data, retaining the reflection intensity and spatial coordinates, and adding BeiDou coordinates and a blockchain timestamp to each point cloud data.

[0072] Then, the LLM large language model is used to parse the relative position descriptions (such as "left shoulder wound") in the text report based on the forensic parameter library, and the relative positions (such as "left shoulder") in the text description are converted into three-dimensional coordinates (x, y, z).

[0073] Specifically, one exemplary method for semantic coordinate transformation is to locate the spatial coordinates corresponding to the "left shoulder" based on a human skeletal model (such as SMPL).

[0074] It should be noted that descriptive language (such as "left shoulder") may cause coordinate deviations due to different perspectives;

[0075] Context-aware LLM large language parsing can be introduced to correct coordinates by combining the suspect's pose in the laser point cloud;

[0076] This data is then fused with the calibrated laser point cloud and the surveillance video stream to generate a spatiotemporally aligned original evidence package.

[0077] It should be noted that by fusing multimodal data from the field with the dual benchmarks of BeiDou and blockchain, the problem of spatiotemporal discontinuity in cross-modal data can be solved.

[0078] The methods for generating evidence fusion packages include:

[0079] Based on the original evidence package, numerical and semantic conflicts between multimodal data are detected.

[0080] The DBSCAN clustering algorithm is used to analyze the suspect's trajectory in the video frame. The suspect's movement trajectory in the surveillance video stream (such as the ventilation duct on the east side) is compared with the coordinates of the same timestamp in the laser point cloud (such as the north side). Then, the spatial deviation between the two coordinates is calculated (such as by calculating the Euclidean distance formula). If the spatial deviation exceeds the expectation (such as >0.3m), it is determined that a semantic conflict has been identified.

[0081] Extract the numerical fields from the text report and compare them with the geometric measurement results of the laser point cloud. Calculate the ratio of the absolute difference between the numerical fields and the geometric measurement results to the numerical fields to obtain the deviation result. If the deviation result is greater than expected (such as 0.08 or 8%), it is determined that a numerical conflict has been identified.

[0082] For example, suppose the forensic description of the wound depth in the text report is 3.0cm, and the geometric measurement result of the laser point cloud is 2.7cm. The absolute difference between the two is 0.3cm. The ratio of this difference to 3.0cm is calculated to obtain a deviation of 0.1. The deviation rate is 10%, which is greater than the expected 8%, and is marked as a numerical conflict.

[0083] For numerical conflicts, an adaptive arbitration factor is generated based on the forensic parameter library. If the numerical conflict meets the expectations, the adaptive arbitration factor is accepted for correction; if the numerical conflict does not meet the expectations, the dynamic correction process is initiated.

[0084] For example, the adaptive arbitration factor generation process is as follows:

[0085] Assuming the wound is located in the liver and was caused by a stab wound from a single-edged dagger, the pre-set forensic parameter library is called to query the trauma parameter table and identify the matching parameter table (e.g., "Trauma Type": "Stab Wound", "Standard Error Range": "±10%", "Knife Type": "Single-edged Dagger", "Tissue Type": "Liver"). The standard error range is found to be ±10%.

[0086] The average of the numerical fields in the text report and the geometric measurement values ​​of the laser point cloud (e.g., 2.85 cm) is taken as the adaptive arbitration factor;

[0087] If the deviation rate is 10%, which is within the standard error range, then the adaptive arbitration factor will be used as the corrected numerical field.

[0088] If the deviation rate is greater than the standard error range, the BeiDou-Blockchain dual-reference anchor time-space coordinates are used to recalibrate the cross-modal data. If the deviation rate is still greater than the standard error range, an advanced arbitration algorithm is applied. For example, the cross-modal data is sorted by confidence level and the median is taken as the adaptive arbitration factor. Alternatively, Bayesian correction is used to calculate the posterior probability based on the prior probability (e.g., "single-edged dagger wound depth 3.0cm") and the current observation value (e.g., laser 2.7cm), and the most likely value is output as the adaptive arbitration factor.

[0089] Regarding semantic conflicts, the semantic conflict arbitration rule is defined as prioritizing the use of standard values ​​from the forensic parameter database. When standard values ​​are missing, laser point cloud coordinates are used (because of their higher spatial accuracy). The reason is that in judicial practice, forensic parameter databases (such as standard wound depth) are more authoritative than point cloud coordinates. When there are no corresponding standard values ​​in the forensic parameter database, laser point cloud coordinates are used.

[0090] For example, the revised trajectory shows the suspect moving to the north ventilation duct;

[0091] If there are no numerical conflicts, a federated weighted confidence mechanism is constructed, and the reliability (i.e. confidence weight) of the text report and the laser point cloud is dynamically allocated through a federated weighted algorithm to generate fused data;

[0092] Example confidence weight allocation results: laser point cloud weight 0.67 (accuracy 0.1mm, reliable geometric measurement), text report weight 0.33 (relies on LLM large language model parsing, which has semantic ambiguity);

[0093] Assuming the geometric measurement of the wound depth in the laser point cloud is 2.85cm, and the wound depth is described as 3.0cm in the text report, the weighted fusion is 0.67×2.85+0.33×3.0≈2.89cm, and 2.89cm is taken as the final numerical field.

[0094] Combined with conflict resolution records and reliability allocation results, a physical evidence fusion package is generated (containing corrected data, arbitration logs of conflict resolution records, confidence weight allocation results, and original evidence package).

[0095] The methods for constructing a resource slicing strategy to divide four-dimensional resource levels include:

[0096] Based on the physical evidence fusion package, a resource slicing strategy is constructed to quantify spatiotemporal density, physical evidence value, behavioral complexity, and environmental interference as four-dimensional resource characteristics, and these are further divided into different levels to obtain four-dimensional resource levels.

[0097] Specifically, the resource slicing strategy quantifies the key features in the evidence fusion package, defines four dimensions and calculation methods, and divides them into different levels (such as dividing the corresponding thresholds of levels A, B, and C by threshold segmentation).

[0098] The calculation method for spatiotemporal density is defined as the proportion of multimodal data (such as video frame count, point cloud density, and text field count) within a certain time period (e.g., 5 minutes) before the incident. For example, in a certain bombing case, the spatiotemporal density is 92%, which exceeds the threshold of Grade A (e.g., >90%), and the grade is classified as Grade A (high).

[0099] The value of physical evidence is calculated as follows: the number of key physical evidence (such as explosive residue, bloodstains, fingerprints, etc.); for example, in a certain bombing case, there are 3 key pieces of evidence, which exceeds the threshold of Grade A (e.g., ≥3), and the grade is classified as Grade A (high).

[0100] The calculation method for behavioral complexity is defined as: the number of categories of suspect actions (such as climbing, running, hiding, etc.); for example, in a bombing case, the suspect's actions have 5 categories, which fall within the range of level B (such as categories 3-5), and the level is classified as level B (medium).

[0101] The calculation method for environmental interference is defined as: weather impact coefficient (e.g., heavy rain = 0.8, cloudy = 0.5, sunny = 0.2); for example, in a certain explosion case, the weather impact coefficient is 0.8, which falls within the range of level B (e.g., 0.5-0.8), and the level is classified as level B (medium).

[0102] It should be noted that the classification rules are as follows: Grade A (High) → High resource requirement; Grade B (Medium) → Moderate resource requirement; Grade C (Low) → Low resource requirement;

[0103] The four-dimensional resource level is mapped to a numerical value, and the GPU resource preemption priority is obtained by combining the four-dimensional resource characteristics.

[0104] Specifically, this involves assigning a weight to each of the four-dimensional resource features. The logic for weight assignment is as follows:

[0105] Spatiotemporal density (e.g., 40%): High spatiotemporal density scenarios (e.g., bombings) require rapid processing of massive amounts of data;

[0106] Value of physical evidence (e.g., 30%): High-value physical evidence (e.g., explosive residue) directly affects the nature of the case;

[0107] Behavioral complexity (e.g., 20%): Complex behaviors require more computational resources to analyze motion trajectories;

[0108] Environmental interference (e.g., 10%): Severe environments (e.g., heavy rain) require additional correction for data noise;

[0109] Then, the corresponding levels of the four-dimensional resource levels are mapped to values ​​in the range of [0, 1], and then weighted fusion is performed to obtain a fusion value; different priorities are divided according to the fusion value by threshold segmentation method, and then the corresponding priorities are matched according to the obtained fusion value;

[0110] Taking a certain bombing case as an example, the spatiotemporal density is: Level A → 1.0; the value of physical evidence is: Level A → 1.0; the complexity of the behavior is: Level B → 0.8; and the environmental interference is: Level B → 0.8.

[0111] Fusion value = 0.4 × 1.0 + 0.3 × 1.0 + 0.2 × 0.8 + 0.1 × 0.8 = 0.94;

[0112] Assuming the priority is divided into [0.75, 0.9], the priority is divided into low priority (<0.75), medium priority (≥0.75 and ≤0.9), and high priority (>0.9).

[0113] Based on the fusion value (0.94) of a certain bombing case, the GPU resource priority of the bombing case is high priority (>0.9).

[0114] GPU resources are dynamically allocated based on GPU resource preemption priority (different amounts of resources are allocated proportionally according to priority, including the number of GPU cores, video memory, and execution time. The allocation logic is: more GPU resources (number of cores and video memory) are allocated to high-priority cases and the execution time is shortened), generating a resource slicing strategy table (containing four-dimensional resource levels and corresponding values, GPU resource preemption priority, and dynamically allocated GPU resources).

[0115] Constructing a simplified set of spatiotemporal trajectories; synchronously triggering case simulations and generating resource arbitration logs include the following methods:

[0116] Based on the evidence fusion package and the optimized resource slicing strategy table, the language description in the text report (such as "the suspect climbed the ventilation duct on the east side and then turned left") is converted into a structured action chain ("climbing" → "ventilation duct" → "east side" → "moving" → "turning left"), and the action attributes are marked, including action type, location, and direction;

[0117] Example of annotated action chains: "Action": "Climb", "Location": "Ventilation duct", "Direction": "East"; "Action": "Move", "Direction": "Turn left";

[0118] Based on the labeled action chain, the key coordinates (such as hands, feet, limb joints, etc.) of the subject (such as the suspect) are mapped into the action chain to generate a topological structure about skeletal movements, which serves as the skeletal movement topological chain.

[0119] The three-dimensional space of the laser point cloud is voxelized and mapped to the voxel index using BeiDou coordinates. Then, through time slicing, a simplified set of spatiotemporal trajectories is constructed by combining the skeletal motion topology chain.

[0120] Specifically, the laser point cloud is divided into voxels of 1m³ (resolution of 0.1m), and then the BeiDou coordinates are mapped to the voxel index (e.g., [120.35,30.27,2.1]→voxel cell_09:32:15.230).

[0121] Then, set a time interval (e.g., 1ms, 10ms), slice according to the time interval, record the timestamp of each voxel, and combine it with the skeletal motion topology chain to construct a simplified set of spatiotemporal trajectories.

[0122] It should be noted that this spatiotemporal trajectory condensation set is achieved by voxelizing spatial and temporal slices, compressing continuous multimodal trajectory data (such as laser point clouds, surveillance videos, and text reports) into discrete spatiotemporal units (voxels), and embedding semantic information (action chains) and causal relationships (skeleton action topology chains) to form a structured spatiotemporal data model. Its essence is a high-dimensional compressed representation of spatiotemporal data, which includes four dimensions: spatial dimension, temporal dimension, semantic dimension, and causal dimension.

[0123] This compressed spatiotemporal trajectory set achieves structured storage and reasoning analysis of multimodal trajectory data through collaborative modeling of four-dimensional characteristics (space, time, semantics, and causality). Its core value lies in reducing redundant storage, preserving behavioral semantics, and providing basic data support for subsequent causal reasoning and resource scheduling.

[0124] It can support efficient behavior pattern recognition and causal reasoning; combined with GPU resource preemption priority, it can dynamically allocate computing resources and accelerate trajectory extrapolation.

[0125] The initial case simulation is conducted based on the GPU resource preemption priority. The purpose of this simulation is to investigate the resource usage during each case simulation, synchronously record GPU resource preemption events during the case simulation, and generate a resource arbitration log (including case ID, preemption time, and GPU utilization rate).

[0126] The methods for constructing a set of physical deduction rules include:

[0127] Based on the simplified set of spatiotemporal trajectories, blood rheological parameters are loaded, and the corresponding trauma parameter table is obtained by querying the forensic parameter library.

[0128] Specifically, the rheological parameters of bloodstains are based on the Navier-Stokes equations. The Navier-Stokes solver is called to define blood viscosity and density. Then, the flow behavior of bloodstains in the spatiotemporal trajectory is simulated based on the Navier-Stokes equations. Combined with the simplified set of spatiotemporal trajectories (containing bloodstain locations and timestamps), the bloodstain diffusion path can be simulated.

[0129] Embedding the trauma parameter table from the forensic parameter library into the physical deduction model can be used to constrain the boundary conditions of the physical deduction, that is, to be used for trauma behavior (such as the wound depth under the action of a knife).

[0130] The trauma parameter table consists of industry-standard parameters pre-set based on forensic theory. Examples of parameter mappings include: Blade type: defines the geometry and cutting force distribution of the blade; Incident angle: calculates the impact of the contact angle between the blade and tissue on wound depth; Tissue type: adjusts the material response model based on the elastic modulus of the internal organs; Standard wound depth (e.g., 4.2±0.3cm): serves as a boundary condition for physical deduction (e.g., maximum wound depth threshold).

[0131] Then, the trauma parameter table is mapped to physical boundary conditions, and blood rheological parameters are integrated to construct a set of physical deduction rules, including fluid equations (Navier-Stokes) + trauma parameters (blade, angle of incidence, tissue type, standard wound depth).

[0132] Used to provide physical constraints for subsequent case simulations, ensuring that the simulation results conform to the laws of biomechanics and forensic medicine.

[0133] The methods for generating physical deduction instruction sets include:

[0134] Match environmental records with the timestamp corresponding to the case from the external environmental parameter database to obtain the risk of impact from rainstorm erosion;

[0135] Based on the set of physical deduction rules and combined with the risk of impact from heavy rain erosion, the area of ​​bloodstain diffusion was corrected.

[0136] Specifically, environmental records corresponding to the time of the case are extracted from the external environmental parameter database, namely the rainfall at the time of the case. The bloodstain diffusion area is adjusted by the rainfall to simulate the washing effect of heavy rain on the bloodstain.

[0137] Based on the rainfall, the scouring factor is calculated and used as the risk of rainstorm scouring impact. The risk of rainstorm scouring impact is used to correct the bloodstain diffusion area obtained by simulation through the physical deduction rule set.

[0138] An exemplary scouring factor formula is as follows: Where Rainfall represents the amount of rainfall (in mm / h), k represents the slope coefficient (by observing the changes in the area of ​​bloodstain diffusion under different amounts of rainfall in experiments, fitting the most suitable parameter, such as 0.15, which means that for every 10 mm / h increase in rainfall, the scouring intensity increases by 15% (such as 75% when it is 50 mm / h)), and the denominator 10 is used to normalize the amount of rainfall to avoid the problem of being too sensitive due to excessively large values;

[0139] The formula adopts the linear form of 1+k×Rainfall to reflect the monotonically increasing effect of rainfall on scour intensity. The formula is essentially a simplified engineering application of the Navier-Stokes equation. Rain scour is essentially the stripping effect of water flow shear force on surface material (bloodstains), and its intensity is positively correlated with the kinetic energy of rainfall.

[0140] The corrected bloodstain diffusion area is obtained by multiplying the scouring factor with the simulated bloodstain diffusion area.

[0141] Example: When the rainfall is 50 mm / h, the scouring factor = 1 + 0.15 × (50 / 10) = 1.75; this means that when the rainfall is 50 mm / h, the area of ​​bloodstain diffusion is corrected to 1.75 times the original area.

[0142] Simultaneously, rainstorm levels are classified based on the risk of erosion from heavy rain, and the wound depth of the standard wound depth is dynamically adjusted according to the rainstorm level; then, environmental parameters (rainfall, rainstorm level) and correction parameters (corrected bloodstain diffusion area, actual wound depth) are integrated to generate a physical deduction instruction set;

[0143] It should be noted that heavy rain may cause swelling of the tissue surface and blurring of the wound edges, resulting in a reduction in the measured wound depth. In this case, the impact of heavy rain on the wound depth measurement needs to be considered, and therefore the wound depth needs to be corrected.

[0144] Specifically, the level of rainstorm can be classified based on historical data using methods such as machine learning;

[0145] Rainstorm levels are usually related to rainfall intensity (e.g., mm / h) or duration. Mapping them to a wound depth correction factor is a simplified modeling approach.

[0146] The method for dynamically adjusting the wound depth is: actual wound depth = standard wound depth × (1 - a × rainstorm level).

[0147] Where 'a' represents an empirical coefficient (e.g., 0.05). Through simulation experiments of knife penetration under different rainstorm levels, it was found that the wound depth decreases linearly with the increase of rainstorm level. Therefore, the most suitable empirical coefficient can be fitted by statistically analyzing the difference between the measured wound depth in rainstorm cases and the standard wound depth.

[0148] Assuming a rainfall of 50 mm / h corresponds to a rainstorm level of 3, a standard wound depth of 4.2 cm, and an empirical coefficient a of 0.05, the actual wound depth after the rainstorm is 4.2 × (1 - 0.05 × 3) = 3.57 cm;

[0149] It should be noted that the goal of this dynamic adjustment of wound depth is to simulate the physical impact of heavy rain on the depth of trauma, rather than to directly correct the measured values ​​in actual cases.

[0150] Standard wound depth is a theoretical value in forensic medicine based on the type of dagger, angle of incidence, and tissue type (e.g., "the standard wound depth of a single-edged dagger piercing the liver is 4.2 cm"), representing the ideal state without environmental interference.

[0151] The actual wound depth is a simulation result under environmental interference (such as heavy rain) and is used to extrapolate case scenarios (e.g., "the wound depth should be 3.57cm under heavy rain conditions").

[0152] Therefore, it is assumed here that the rainstorm is an external disturbance factor to the wound depth, and its role is to correct the benchmark value of standard wound depth, rather than to correct the measured value in actual cases.

[0153] The methods for generating 3D simulation animation streams include:

[0154] Based on the GPU utilization rate in the resource arbitration log and the GPU resource preemption priority of each case, the computation scheduling strategy for case simulation is dynamically adjusted.

[0155] Specifically, in the process of case deduction, it is necessary to complete the human body movements. Therefore, the skeletal movement topology chain can be used as input, and the complete skeletal trajectory can be output through the RPM rolling prediction model.

[0156] During the inference process of the pre-trained RPM rolling prediction model, the batch size of the model inference is adjusted in real time according to the GPU utilization rate of the resource arbitration log to avoid resource waste or conflicts.

[0157] During the inference process of the pre-trained GAN model, if the current GPU utilization rate is lower than the preset threshold (e.g., 80%), the GAN model is allowed to run at full speed; otherwise, the generation resolution is reduced (e.g., from 1080p to 720p) according to the GPU resource preemption priority in the resource arbitration log.

[0158] According to the adjusted computational scheduling strategy, the skeletal motion topology chain is completed by the RPM rolling prediction model. After completion, it is used as a complete skeletal trajectory and combined with the physical deduction instruction set. The action instructions and physical deduction set are input into the GAN model to generate frame-level behavior descriptions, and then output a three-dimensional behavior deduction animation stream.

[0159] Synchronously record the GPU usage status and computing scheduling strategy of the current case (including batch size and resolution adjustment), and update the resource arbitration log (the updated resource arbitration log includes the GPU utilization and computing scheduling strategy record of the current case).

[0160] The methods for generating crime logic analysis reports include:

[0161] Based on the 3D behavior deduction animation flow, combined with the physical deduction instruction set and the spatiotemporal trajectory simplification set, physical evidence is associated with action behavior to construct a causal logic chain of criminal behavior, and then the causal chain conclusion is obtained.

[0162] Specifically, Prolog logic rules are constructed using the Prolog logic language to define causal relationships, such as: motive (robbery), cause: physical evidence (pry marks on the safe, knife wounds), behavior (lock picking, bayonet), time (two hours before the case was discovered).

[0163] Parameter sources: physical evidence is extracted from the concise spatiotemporal trajectory (such as the location of the safe and pry marks, knife wounds), behavior is identified from the 3D behavior deduction animation stream (such as identification through frame-level action description), and time can be extracted from the case timestamp or the metadata of the 3D behavior deduction animation stream;

[0164] Based on the constructed Prolog logic rules, the Prolog interpreter is invoked, with the constructed Prolog logic rules and case data as input, and the causal chain conclusion is output.

[0165] Example: Input:

[0166] Safe: Located in the southeast corner of the crime scene (from the simplified time-space trajectory set); Trauma: Located on the left side of the victim's abdomen (from the simplified time-space trajectory set);

[0167] Lock picking: The action was identified in frame 12 of the animation stream; Bayonet: The action was identified in frame 5 of the 3D behavior deduction animation stream.

[0168] Time: 2 hours before the case was discovered (from case metadata);

[0169] Output:

[0170] The motive (robbery and murder) is established, meaning the criminal motive is logically verified;

[0171] Based on the causal chain conclusion, four-dimensional spatiotemporal coordinates are bound to the causal logic chain to generate spatiotemporal binding labels, and then integrated to generate a crime logic analysis report (including causal chain conclusion, spatiotemporal binding labels, and physical evidence behavior correlation data).

[0172] Specifically, based on the BeiDou coordinates mapped to the voxel index within the simplified set of spatiotemporal trajectories (e.g., [120.35,30.27,2.1] → voxel cell_09:32:15.230), these coordinates are used as four-dimensional spatiotemporal coordinates. Combined with the first frame of the three-dimensional behavior deduction animation stream (containing the initial behavior state (e.g., suspect's position, action start point)) as input, a unique identifier is generated using the SHA-256 algorithm, serving as a spatiotemporal binding label: label = hash algorithm(four-dimensional spatiotemporal coordinates || first frame of animation stream);

[0173] Example: tag=hash("4D@09:32:15.230"||"Frame 1: Right hand gripping the edge of the pipe")→"7d3fe4a6b8f1c9e...";

[0174] The purpose of this binding is to link hash tags to causal chain conclusions, forming a spatiotemporal anchor point to ensure the immutability of the crime logic analysis report.

[0175] Methods for optimizing and updating forensic parameter databases and resource slide strategies include:

[0176] The system identifies sensitive fields in the 3D simulation animation stream and crime logic analysis report, and uses hot-loading desensitization rules to desensitize and replace these sensitive fields. The aim is to protect sensitive information (such as DNA and victim identity) during the case simulation iteration process.

[0177] Specifically, use regular expressions or pattern matching tools to identify sensitive fields (such as victim DNA).

[0178] Then, the desensitization rule is defined as replacing the middle part of the sensitive field with *, while retaining the unique identifier (e.g., 1623); Sensitive field: Victim DNA: ATCG-1623..., after desensitization: A**-****-1623;

[0179] The hot-loading mechanism injects the desensitized data into the model training process of case simulation in real time, thus preventing the original sensitive data from entering the local computing environment.

[0180] A judicial credibility mechanism is established to quantify the degree of matching between the deduction results and the actual evidence in the physical evidence fusion package, thereby obtaining the credibility of the case deduction results;

[0181] Specifically, the judicial credibility mechanism defines credibility indicators, then quantifies the credibility indicators, and then quantifies the degree of matching between the inference results and the actual evidence in the physical evidence fusion package, thereby obtaining the credibility of the case inference results.

[0182] For example:

[0183] The credibility indexes are defined as the error in depth of creation, the reasonableness score of behavior, and the environmental response score.

[0184] Blow depth error: The relative error between the physical deduction of the actual blow depth and the actual measurement value (absolute value, range [0,1]);

[0185] Example: If the physical deduction of the actual wound depth is 3.57cm, and the actual measurement is 3.42cm, the error is |3.57-3.42| / 3.42≈0.0438;

[0186] Reasonableness score for behavior: a score based on the reasonableness of actions in the 3D simulation animation stream (range [0,1]);

[0187] Example: The frame-level description confidence level for the lock-picking action is 0.88;

[0188] Environmental response score: The matching degree score between environmental parameters (such as rainstorm level) in the physical deduction instruction set and the behavior of the animation flow (range [0,1]);

[0189] Example: The match between the rainstorm level 3 and the bloodstain spread area correction value is 0.91;

[0190] Then, weights are assigned to these credibility indicators, and a weighted fusion calculation is performed to obtain the final credibility score.

[0191] The credibility level is divided by the threshold segmentation method, and the corresponding credibility level is obtained according to the calculated credibility, which is used as the credibility of the case deduction result.

[0192] For example, a confidence level ≥ 0.9 is considered high confidence, 0.85 ≤ confidence level < 0.9 is considered medium confidence, and confidence level < 0.85 is considered low confidence.

[0193] Set up an update feedback mechanism: High credibility: Directly include in the forensic parameter database update; Medium credibility: Mark as a case pending review and trigger manual review process; Low credibility: Backtrack the physical deduction instruction set and animation flow to adjust model parameters;

[0194] Based on the credibility results, a federated optimization algorithm is used to optimize the forensic parameter library and resource scheduling strategy through distributed collaboration.

[0195] Specifically, based on the credibility results, the forensic parameter library is updated locally and synchronized to the central server for global aggregation, resulting in an updated forensic parameter library and an optimized resource slicing strategy;

[0196] For example, a parameter update formula can be set, such as: Forensic parameter library. Knife wound model += 0.01 × gradient;

[0197] The gradient is calculated by backpropagating the confidence level results and the physical deduction error.

[0198] Example: If the current tool trauma model parameters are The gradient is Updated parameters ;

[0199] Each local node only updates its local model and does not share sensitive data (such as anonymized case data).

[0200] Set a synchronization frequency (e.g., 10 cases), and then perform global aggregation based on the synchronization frequency;

[0201] For example, the parameters of the case deduction model are synchronized every 10 cases, and then the global aggregation strategy is to use a weighted average method, where the weights are proportional to the credibility of the case.

[0202] Finally, the output results are an updated forensic parameter library (an incremental package containing parameters such as knife wound models, environmental correction coefficients (slope coefficients of scour factors, empirical coefficients of actual wound depth)) and an optimized resource slicing strategy (a weight table of four-dimensional resource features used to adjust case priority allocation rules).

[0203] Example 2

[0204] Please see Figure 3 As shown, for parts not described in detail in this embodiment, please refer to the description in Embodiment 1. This embodiment provides a virtual reality method for three-dimensional scene reconstruction and investigative deduction in criminal cases, including:

[0205] S1: Integrate cross-modal evidence from the scene, use BeiDou-Blockchain dual-reference anchored spatiotemporal coordinates to detect conflicts in physical evidence; construct an adaptive arbitration factor based on a forensic parameter library to generate a physical evidence fusion package;

[0206] S2: Analyze the physical evidence fusion package, construct a resource slicing strategy to classify four-dimensional resource levels; transform text descriptions into skeletal motion topology chains, construct a simplified set of spatiotemporal trajectories; synchronously trigger case deduction, and generate resource arbitration logs;

[0207] S3: Based on the simplified set of spatiotemporal trajectories and resource arbitration logs, construct a set of physical inference rules; then execute rainstorm erosion correction to generate a set of physical inference instructions;

[0208] S4: Based on the physical deduction instruction set, perform case deduction resource scheduling and generate a three-dimensional deduction animation stream; then construct a judicial causal chain, bind it to a spatiotemporal version, and generate a crime logic analysis report;

[0209] S5: Based on the 3D simulation animation flow and crime logic analysis report, hot-load desensitization rules are applied to identify and replace sensitive fields; the effectiveness of the simulation is quantified, and the forensic parameter library and resource slicing strategy are optimized and updated.

[0210] Example 3

[0211] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases described above.

[0212] Since the electronic device described in this embodiment is used to implement the virtual reality method for three-dimensional scene reconstruction and investigation deduction of criminal cases according to the embodiments of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the virtual reality method for three-dimensional scene reconstruction and investigation deduction of criminal cases described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. As long as those skilled in the art implement the virtual reality method for three-dimensional scene reconstruction and investigation deduction of criminal cases according to the embodiments of this application, the electronic device used is within the scope of protection of this application.

[0213] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0214] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A virtual reality platform for three-dimensional scene reconstruction and investigative simulation in criminal investigation cases, characterized in that, include: Cross-modal evidence fusion unit: Integrates cross-modal evidence from the scene, and uses BeiDou-blockchain dual-reference anchor spatiotemporal coordinates to detect conflicts in physical evidence; An adaptive arbitration factor is constructed based on a forensic parameter database to generate a physical evidence fusion package. Resource Slicing and Trajectory Compression Unit: Parse the physical evidence fusion package and construct a resource slicing strategy to classify four-dimensional resource levels; Transform text descriptions into skeletal motion topology chains to construct a simplified set of spatiotemporal trajectories; Simultaneously trigger case simulations and generate resource arbitration logs; Forensic physical rule injection unit: Based on the simplified set of spatiotemporal trajectories and resource arbitration logs, a set of physical deduction rules is constructed; then, rainstorm erosion correction is performed to generate a set of physical deduction instructions; Criminal investigation behavior synthesis and tracing unit: Based on the physical deduction instruction set, it performs case deduction resource scheduling and generates a three-dimensional deduction animation stream; then it constructs a judicial causal chain, binds it to a spatiotemporal version, and generates a crime logic analysis report; Judicial Trust Iteration Unit: Based on the 3D simulation animation stream and crime logic analysis report, hot-load desensitization rules are applied to identify and replace sensitive fields; Quantify the effectiveness of the extrapolation, optimize and update the forensic parameter library and resource slicing strategy.

2. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 1, characterized in that, The method of detecting physical evidence conflicts by anchoring spatiotemporal coordinates using the BeiDou-Blockchain dual-reference point includes: Integrate on-site laser point clouds, text reports, and surveillance video streams as cross-modal evidence; Map the local coordinates of the laser point cloud to the BeiDou positioning system to establish a global three-dimensional spatial benchmark; align the timestamps of the monitoring video stream with the blockchain timestamps. The relative position description in the text report is loaded from the pre-set forensic parameter library and converted into three-dimensional coordinates. These coordinates are then fused with the calibrated laser point cloud and surveillance video stream to generate a spatiotemporally aligned original evidence package.

3. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 2, characterized in that, The generation methods of the evidence fusion package include: Based on the original evidence package, semantic conflicts are identified by comparing the coordinates of the same timestamp in the surveillance video stream and the laser point cloud. Extract numerical fields from the text report and compare them with the geometric measurement results of the laser point cloud to identify numerical conflicts; For numerical conflicts, an adaptive arbitration factor is constructed based on the forensic parameter library. If the numerical conflict meets the expectations, the adaptive arbitration factor is accepted for correction; if the numerical conflict does not meet the expectations, the dynamic correction process is initiated. For semantic conflicts, the arbitration rule is defined as prioritizing the use of standard values ​​from the forensic parameter library, and using laser point cloud coordinates when standard values ​​are missing; By combining conflict resolution records and reliability allocation results, an evidence fusion package is generated.

4. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 3, characterized in that, The methods for constructing the resource slicing strategy to divide four-dimensional resource levels include: Based on the physical evidence fusion package, a resource slicing strategy is constructed to quantify spatiotemporal density, physical evidence value, behavioral complexity, and environmental interference as four-dimensional resource characteristics, and these are further divided into different levels to obtain four-dimensional resource levels. The four-dimensional resource level is mapped to a numerical value, and the GPU resource preemption priority is obtained by combining the four-dimensional resource characteristics. GPU resources are dynamically allocated based on GPU resource preemption priority, and a resource slicing strategy table is generated.

5. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 4, characterized in that, The construction of a simplified set of spatiotemporal trajectories; The methods for synchronously triggering case simulations and generating resource arbitration logs include: Based on the evidence fusion package and the optimized resource slicing strategy table, the language description in the text report is converted into an action chain, and the action attributes are labeled to generate a skeletal action topology chain. The three-dimensional space of the laser point cloud is voxelized and mapped to the voxel index using BeiDou coordinates. Then, through time slicing, a simplified set of spatiotemporal trajectories is constructed by combining the skeletal motion topology chain. Based on the GPU resource preemption priority, the case is simulated, and GPU resource preemption events during the case simulation are recorded synchronously to generate a resource arbitration log.

6. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 5, characterized in that, The physical deduction rule set is constructed in the following ways: Based on the simplified set of spatiotemporal trajectories, blood rheological parameters are loaded, and the corresponding trauma parameter table is obtained by querying the forensic parameter library. Then, the trauma parameter table is mapped to physical boundary conditions, and blood rheological parameters are fused to construct a set of physical deduction rules.

7. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 6, characterized in that, The methods for generating the physical deduction instruction set include: Match environmental records with the timestamp corresponding to the case from the external environmental parameter database to obtain the risk of impact from rainstorm erosion; Based on the set of physical deduction rules and combined with the risk of impact from heavy rain erosion, the area of ​​bloodstain diffusion was corrected. At the same time, rainstorm levels are classified according to the risk of erosion from rainstorms, and the depth of damage is dynamically adjusted according to the rainstorm level; then, a set of physical deduction instructions is generated.

8. The virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 7, characterized in that, The generation methods of the 3D simulation animation stream include: Based on the GPU utilization rate in the resource arbitration log and the GPU resource preemption priority of each case, the computation scheduling strategy for case simulation is dynamically adjusted. Based on the adjusted computation scheduling strategy, the skeletal motion topology chain is completed and used as motion instructions. Combined with the physical deduction instruction set, a frame-level behavior description is generated, thus obtaining a 3D behavior deduction animation stream. Synchronously record the GPU usage status and computing scheduling strategy for the current case, and update the resource arbitration log.

9. A virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 8, characterized in that, The methods for generating the crime logic analysis report include: Based on the 3D behavior deduction animation flow, combined with the physical deduction instruction set and the spatiotemporal trajectory simplification set, physical evidence is associated with action behavior to construct a causal logic chain of criminal behavior, and then the causal chain conclusion is obtained. Based on the causal chain conclusion, four-dimensional spatiotemporal coordinates are bound to the causal logic chain to generate spatiotemporal binding tags, which are then integrated to generate a crime logic analysis report.

10. A virtual reality platform for three-dimensional scene reconstruction and investigation simulation of criminal cases according to claim 9, characterized in that, The methods for optimizing and updating the forensic parameter database and resource slice strategy include: Identify sensitive fields in the 3D simulation animation stream and crime logic analysis report, and hot-load the desensitization rules to desensitize and replace the sensitive fields; A judicial credibility mechanism is established to quantify the degree of matching between the deduction results and the actual evidence in the physical evidence fusion package, thereby obtaining the credibility of the case deduction results; Based on the reliability results, the forensic parameter library is updated locally and synchronized to the central server for global aggregation, resulting in an updated forensic parameter library and an optimized resource slicing strategy.

Images