A method, medium and system for dynamically generating a roaming path in a three-dimensional scene

Abstract

This invention provides a method, medium, and system for dynamically generating roaming paths in a 3D scene, belonging to the field of 3D technology. This invention performs lossless dimensionality reduction and alignment of multi-source heterogeneous 3D data through pre-orthogonal decoupling and dynamic lightweight topology. It utilizes a learnable mask dynamic sparse perception mechanism and local activation operators to achieve adaptive sparse forward computation of the view frustum. The roamer pose and the spatiotemporal trajectory of dynamic obstacles are input into a conditional random field depth-extended spatiotemporal graph evolution model, and a composite energy function drives asynchronous parallel pipeline multi-channel graph inference to generate multi-track obstacle avoidance roaming curves. A smooth 3D roaming path point set tensor is output. Finally, global data lock waiting is eliminated through asynchronous lightweight local subgraph AI parallel inference. This solves the technical problem of difficulty in real-time fusion of multi-source heterogeneous data and simultaneous high-quality global path planning in a multi-dynamic obstacle environment during the dynamic generation of 3D roaming paths.

Description

Technical Field

[0001] This invention belongs to the field of 3D technology, and specifically relates to a method, medium, and system for dynamically generating roaming paths in a 3D scene. Background Technology

[0002] Dynamic generation technology for 3D scene roaming paths is widely used in virtual reality, digital twins, autonomous driving simulation, and industrial inspection. This requires the system to simultaneously plan a smooth roaming path with obstacle avoidance capabilities while receiving heterogeneous data from multiple sources, such as 3D meshes and point clouds. Current mainstream methods typically employ path planning algorithms based on occupancy grids or potential fields, combined with traditional convolutional neural networks to extract scene features, and then generate the path result using a full-map synchronous inference method. In existing technologies, belief propagation (BP) is a message-passing algorithm used to effectively perform this type of inference. However, in existing technologies, when multiple dynamic obstacles exist simultaneously in a 3D scene and the input data is a mixed streaming data of high-dimensional meshes and point clouds, the traditional full-map synchronous inference structure requires waiting for the global data lock to be released before belief propagation can be performed. Furthermore, the quadratic computational complexity of the fully connected attention mechanism increases dramatically with the scene size. This leads to a cumulative effect between the fusion delay of high-dimensional multi-source heterogeneous data and the computational overhead of global path inference, making it impossible for the system to simultaneously complete multi-source data fusion and global obstacle avoidance path generation under real-time constraints. In other words, existing technologies face the technical challenge of real-time fusion of multi-source heterogeneous data during the dynamic generation of 3D scene roaming paths, and the simultaneous achievement of high-quality global path planning in environments with multiple dynamic obstacles. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for dynamically generating roaming paths in a 3D scene, which can solve the technical problem in the prior art that it is difficult to fuse multi-source heterogeneous data in real time and achieve high-quality global path planning synchronously in a multi-dynamic obstacle environment during the dynamic generation of 3D scene roaming paths.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for dynamically generating a roaming path in a three-dimensional scene, comprising the following steps:

[0005] Pre-orthogonal decoupling is performed on the mesh and point cloud data flowing into the 3D scene, and the high-dimensional mesh features are mapped to the low-dimensional latent space using a dynamic lightweight topology structure, thus completing the topology alignment and lossless dimensionality reduction of multi-source heterogeneous data.

[0006] The rendering features within the current view frustum are extracted, and a learnable mask dynamic sparse perception mechanism is used to retain only the core features of the view frustum. Combined with dynamic reconstruction using local activation operators, instantaneous adaptive jumping of the forward channel of path prediction is achieved.

[0007] The current pose of the 3D rover and the spatiotemporal trajectory sequence of dynamic obstacles are input into the conditional random field depth-extended spatiotemporal graph evolution model. The composite energy function driven by the state entropy increase dynamic constraint function and the group cooperative collision avoidance behavior function is used to infer the multi-track obstacle avoidance rover curve in a multi-channel asynchronous parallel pipeline.

[0008] The dynamic obstacle trajectory sequence tensor and the current state features of the roamer are input into the linearized axial separation attention model of the group collaborative collision avoidance behavior. The gated route weights are adjusted in real time by the particle swarm optimization sparse activation route adaptive update algorithm, and the output is a smooth three-dimensional roaming path point set tensor.

[0009] By using a long-span residual dense jump matrix to transmit historical key viewpoint features across layers, and combining it with an adaptive exponential decay sliding window to lock long-distance spatial features, the global topological coherence of long-distance paths is maintained.

[0010] The three-dimensional spatial topology is decoupled into asynchronous lightweight local subgraphs under a multi-dimensional network transmission and fusion mechanism. Each subgraph thread independently infers through an asynchronous parallel pipeline with multiple channels, eliminating thread waiting caused by global data locks and outputting the final roaming path.

[0011] The aforementioned pre-orthogonal decoupling refers to performing orthogonal basis decomposition on the high-dimensional grid features at the boundary of the streaming data stream, removing the linear correlation between each dimension, making the feature components independent of each other, thereby avoiding redundant calculations in subsequent mapping.

[0012] The dynamic lightweight topology refers to a graph structure constructed based on the principle of spatial self-organization and evolution. The number of nodes and edges and their connection relationships are adjusted in real time according to the spatial distribution of the input data, and the topology scale is automatically reduced under different load conditions. The dimensional range of the low-dimensional potential space is an extremely low-dimensional feature representation space within the low-dimensional ratio threshold range of the original feature dimensions. The low-dimensional ratio threshold is determined by conducting joint experiments on reconstruction error and computational delay on multiple sets of three-dimensional scene data of different scales and performing Pareto optimality analysis on the experimental data.

[0013] The learnable mask dynamic sparse perception mechanism refers to configuring a set of binary mask weights that can be updated with training for each feature channel inside the neural network, and activating or masking the corresponding channel in real time according to the geometric changes of the view frustum during inference; the local activation operator refers to an operator that performs nonlinear activation operations only in the local spatial neighborhood of the feature map; the instantaneous adaptive jump of the path prediction forward channel refers to the fact that when the rendered features in the view frustum change abruptly, the forward propagation process skips the network layers that are unrelated to the current features and directly routes the activation results to the subsequent output layer.

[0014] The specific structure of the Conditional Random Field Deep Extended Spatiotemporal Graph Evolution Model is as follows: The input layer represents the 3D roamer pose sequence and the spatiotemporal trajectory of dynamic obstacles as nodes and directed edges of a graph, and constructs a full graph representation through a topological graph sparse adjacency transmission mechanism; the energy function is composed of a state entropy increase dynamic constraint function and a group cooperative collision avoidance behavior function, and is substituted into the multi-level potential function for calculation; the information flow direction within the network is conditionally controlled by the cascaded gating of the potential function; the model adopts an asynchronous parallel pipeline multi-channel structure, and each sub-layer independently performs belief propagation iteration; local asymmetric gating loop units are used inside each channel to ensure the real-time performance of forward computation; through the deep cascaded correction of the multi-level potential function, the model outputs the optimal spatiotemporal roaming path after the global energy function converges.

[0015] The state entropy increase dynamic constraint function is used to quantify the uncertainty of obstacle trajectory distribution. The higher the entropy value, the stronger the constraint, forcing path points to converge to low-entropy regions. The group cooperative collision avoidance behavior function models the repulsive force between multiple agents as a potential field function. The intensity of the repulsive force is jointly determined by the relative distance and relative velocity. The relative distance threshold is determined by repeatedly conducting obstacle avoidance success rate experiments in obstacle scenarios with different densities and performing piecewise regression analysis on the experimental data. The number of belief propagation iterations is determined by multiple rounds of comparative experiments on convergence speed and computation delay.

[0016] The training dataset establishment steps for the Conditional Random Field Deep Extension Spatiotemporal Graph Evolution Model include: collecting 3D scene simulation data containing randomly moving obstacles of different densities, recording the rover pose sequence, obstacle spatiotemporal trajectory, and corresponding collision-free reference path, constructing a triplet dataset, and augmenting the data by applying random perturbations to the existing trajectories for any deficiencies; performing graph structuring preprocessing on the original trajectory data under the topological graph sparse adjacency propagation mechanism to generate node feature matrices and adjacency matrix pairs; and training steps include: using the composite energy function as the basic term of the loss function, combined with the path smoothness regularization term, and using an asynchronous parallel gradient accumulation strategy to update the network parameters of each subgraph channel, triggering early stopping when the energy function on the validation set does not decrease for multiple consecutive rounds.

[0017] The specific structure of the linearized axial separation attention model for collaborative collision avoidance behavior is as follows: the input layer projects multi-source heterogeneous inputs to a unified low-dimensional latent space through a shared latent space low-dimensional alignment operator; the underlying core adopts a linear attention approximation mechanism and an axial separation attention mechanism to decouple the traditional fully connected attention matrix into one-dimensional attention calculation operators along three single axes: X-axis, Y-axis, and Z-axis, with each operator introducing a local sliding window; the behavioral repulsion function of collaborative collision avoidance behavior is embedded in the linear approximation operator as a weight scaling factor; the information flow is controlled in real time by cascaded gated attention; a long-span residual dense jump matrix spans multiple axial attention layers; and the output layer generates a smooth three-dimensional roaming path point set tensor.

[0018] The particle swarm optimization sparse activation route adaptive update algorithm is nested within the gated routing mechanism of the linearized axial separation attention model of the group's cooperative collision avoidance behavior. When initializing the particle swarm, the position of each particle represents a sampling point in the gated routing weight parameter space, and the velocity represents the parameter update step size. The algorithm calculates the real-time prediction error rate under the current route output and combines the prediction error rate with the path smoothness index to form a fitness function. In each iteration, each particle adjusts its update step size based on its own historical best fitness and the group's global best fitness. The updated particle position is directly mapped to the real-time selection weight of the gated route.

[0019] The sparse activation routing dynamic adjustment function is based on the mean value of the repulsive force field in the current frame. Path smoothness index Compared with the current prediction error rate Calculate the comprehensive routing pressure index using three data points. The calculation method is as follows ,in , , The weighting coefficients and ;when Different adjustment strategies are used for the sparse activation threshold parameter of the gated routing layer when the ranges are different: when When the stress threshold is below the low-pressure threshold, the sparse activation threshold is adjusted to the high-sparseness range; when When within the medium pressure threshold range, the sparse activation threshold remains in the medium sparsity range; when When the pressure threshold is above the high pressure threshold, the sparse activation threshold is reduced to the low sparsity range and all axial attention operators are activated.

[0020] The long-span residual dense jump matrix refers to a dense cross-layer connection weight matrix established between multiple non-adjacent layers of the network, where each cross-layer connection path carries an independent learnable scaling coefficient. The adaptive exponential decay sliding window refers to a sliding time window mechanism that retains historical key viewpoint features by weighting them according to an exponential decay coefficient. The weight of historical features further away from the current frame is lower. The decay coefficient is determined through multiple rounds of experiments on the topological coherence of long-distance paths under different decay coefficients and statistical analysis of trajectory continuity indicators.

[0021] The asynchronous lightweight local subgraph refers to the independent subgraph unit obtained by dividing the three-dimensional spatial topology into spatial partitions. Each subgraph only maintains the nodes and edges within its own partition, and the boundary features of each subgraph are synchronized through an asynchronous message passing mechanism. The asynchronous parallel pipeline multi-channel refers to a parallel scheduling structure in which multiple subgraph inference channels are executed in an interleaved manner without waiting for each other to complete. Each channel independently holds a local buffer, eliminating thread mutual exclusion waiting caused by global data locks.

[0022] The conditional random field deep extended spatiotemporal graph evolution model extends the single-layer potential function of the traditional graph-theoretic Markov random field into a multi-level deep cascaded potential function, enabling the model to refine the spatiotemporal interaction probability between obstacles and roamers layer by layer on the graph structure. The locally asymmetric gated recurrent unit refers to an improved gated recurrent unit in which the input gate and forget gate weight matrices are not shared. Through asymmetric weight design, the unit has different selectivity for input features and historical states.

[0023] The shared latent space low-dimensional alignment operator refers to an alignment operation operator that applies independent linear projections to multi-source heterogeneous inputs and then stitches the projection results into the same low-dimensional latent space; the linear attention approximation mechanism refers to an approximation calculation method that uses kernel function decomposition to reduce the quadratic computational complexity of standard attention to linear level; the rigid obstacle avoidance property refers to the property that the path point set satisfies the constraint of maintaining a safe distance from obstacles at any time, and the safe distance is determined by conducting collision rate statistical experiments under different combinations of obstacle size and movement speed.

[0024] The low-dimensional latent space has a dimension ranging from 2% to 8% of the original feature dimension; the relative distance threshold ranges from 0.5 to 3.0 m; the number of belief propagation iterations ranges from 3 to 8; the number of particles ranges from 10 to 50, and the number of iterations per round ranges from 5 to 20; the weight coefficients... , , The ranges are all 0.2–0.6; the low pressure threshold is 0.3, and the high pressure threshold is 0.6; the attenuation coefficient ranges from 0.85 to 0.99; the safe distance ranges from 0.3 to 2.0 m; the local sliding window size ranges from 8 to 32 feature units; the weighting coefficient ranges from 0.3 to 0.7; and the axial attention layer learning rate ranges from [missing value]. ~ The learning rate range of the gated routing layer is ~ .

[0025] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for dynamically generating a roaming path in a three-dimensional scene.

[0026] A third aspect of the present invention provides a dynamic path generation system for a three-dimensional scene, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0027] This invention reduces and aligns high-dimensional, multi-source heterogeneous data to a low-dimensional latent space through pre-orthogonal decoupling and dynamic lightweight topology, eliminating redundant computations at the source and enabling all subsequent modules to run on extremely low-dimensional features. Utilizing a learnable mask dynamic sparse perception mechanism and local activation operators, computation is performed only on core features when the view frustum changes, avoiding the redundant overhead of global activation. Employing a conditional random field deep-extended spatiotemporal graph evolution model and an asynchronous parallel pipeline multi-channel structure eliminates thread waiting caused by global data locks, allowing each subgraph to infer independently before lightweight boundary alignment, enabling the global energy function to converge in a very short time in multi-dynamic obstacle scenarios. Introducing a linear attention approximation mechanism and an axial separation attention mechanism decouples quadratic attention computation into three single-axis linear operators, combined with a particle swarm optimization sparse activation route adaptive update algorithm to adjust the gated route weights in real time, further compressing the computational resource consumption of real-time inference. In summary, this invention solves the technical problem mentioned in the background art of the difficulty in real-time fusion of multi-source heterogeneous data and the synchronous realization of high-quality global path planning in a multi-dynamic obstacle environment during the dynamic generation of 3D scene roaming paths. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This diagram illustrates the data flow for pre-orthogonal decoupling and dimensionality reduction alignment of the low-dimensional latent space.

[0030] Figure 3 This is a schematic diagram of the network structure of the deep extended spatiotemporal graph evolution model of conditional random fields.

[0031] Figure 4 A schematic diagram illustrating the process of generating the path point set tensor for a linearized axial separation attention model that enables collaborative collision avoidance in groups.

[0032] Figure 5 This is a schematic diagram of the feature transmission structure of a long-span residual dense jump matrix and an adaptive exponential decay sliding window.

[0033] Figure 6 This is a schematic diagram of an asynchronous parallel pipeline multi-channel inference scheduling structure.

[0034] Figure 7 This is a map showing the topological coherence of the entire roaming path and the distribution of safe distances between path points in each segment. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0036] like Figure 1 The diagram shown is a flowchart of a method for dynamically generating a roaming path in a 3D scene according to the first aspect of the present invention. This method includes the following steps:

[0037] S01. Perform pre-orthogonal decoupling on the mesh and point cloud data flowing into the 3D scene, and use a dynamic lightweight topology structure to map the high-dimensional mesh features to the low-dimensional potential space to complete the topology alignment and lossless dimensionality reduction of multi-source heterogeneous data.

[0038] S02. Extract the rendering features within the current view frustum, utilize the learnable mask dynamic sparse perception mechanism to retain only the core features of the view frustum, and combine local activation operators for dynamic reconstruction to achieve instantaneous adaptive jumping of the forward channel of path prediction.

[0039] S03. Input the tensor of the current pose of the 3D rover and the spatiotemporal trajectory sequence of dynamic obstacles into the conditional random field depth-extended spatiotemporal graph evolution model. Use the composite energy function driven by the state entropy increase dynamic constraint function and the group cooperative collision avoidance behavior function to infer the multi-track obstacle avoidance rover curve with rigid obstacle avoidance properties in the multi-channel asynchronous parallel pipeline.

[0040] S04. Input the dynamic obstacle trajectory sequence tensor and the current state features of the roamer into the linearized axial separation attention model of the group collaborative collision avoidance behavior. The gating route weights are adjusted in real time by the particle swarm optimization sparse activation route adaptive update algorithm, and the output is a smooth three-dimensional roaming path point set tensor.

[0041] S05. Historical key viewpoint features are transmitted across layers through a long-span residual dense jump matrix, and long-distance spatial features are locked by an adaptive exponential decay sliding window, thus maintaining the global topological coherence of long-distance paths.

[0042] S06. The three-dimensional spatial topology is decoupled into asynchronous lightweight local subgraphs under the multi-dimensional network transmission and fusion mechanism. Each subgraph thread independently infers through multiple channels of asynchronous parallel pipeline, eliminating thread waiting caused by global data locks, and outputting the final roaming path.

[0043] Among them, pre-orthogonal decoupling refers to performing orthogonal basis decomposition on high-dimensional grid features at the boundary of the streaming data stream, removing the linear correlation between dimensions, making the feature components independent of each other, thereby avoiding redundant calculations in subsequent mapping; dynamic lightweight topology refers to a graph structure built based on the principle of spatial self-organization evolution, in which the number of nodes and edges and the connection relationship are adjusted in real time according to the spatial distribution of the input data, and the topology scale is automatically reduced under different load conditions; low-dimensional latent space refers to the extremely low-dimensional feature representation space obtained by compressing high-dimensional grid features through local nonlinear state space mapping. The recommended dimension range is 2% to 8% of the original feature dimension. The specific value is determined by conducting joint experiments on reconstruction error and computation delay on multiple sets of 3D scene data of different scales, and then performing Pareto optimality analysis on the experimental data.

[0044] Among them, the learnable mask dynamic sparse perception mechanism refers to configuring a set of binary mask weights that can be updated with training for each feature channel inside the neural network. During inference, the corresponding channels are activated or masked in real time according to the geometric changes of the view frustum, so that the network only performs calculations on the core feature items. The local activation operator refers to the operator that performs nonlinear activation operations only in the local spatial neighborhood of the feature map, avoiding the redundant computational overhead caused by global activation. The instantaneous adaptive jump of the path prediction forward channel means that when the rendered features in the view frustum change abruptly, the forward propagation process skips the network layers that are not related to the current features and directly routes the activation results to the subsequent output layer.

[0045] The specific structure of the Conditional Random Field Deep Extended Spatiotemporal Graph Evolution Model is as follows: The model input layer represents the 3D roamer pose sequence and the spatiotemporal trajectory of dynamic obstacles as nodes and directed edges of a graph, and constructs a full graph representation through the sparse adjacency transmission mechanism of the topological graph; the energy function is composed of the state entropy increase dynamic constraint function and the group cooperative collision avoidance behavior function, and is directly substituted into the multi-level potential function for calculation; the information flow direction inside the network is conditionally controlled by the cascaded gating of the potential function, and the ultra-lightweight cascaded gating intercepts low potential energy feature branches in advance before heavy edge inference; the model adopts an asynchronous parallel pipeline multi-channel structure, and each sub-layer independently performs belief propagation iteration; local asymmetric gating loop units are used inside each channel to ensure the real-time performance of forward computation; through the deep cascaded correction of the multi-level potential function, the model outputs the optimal spatiotemporal roaming path after the global energy function converges. The state entropy increase dynamic constraint function is used to quantify the uncertainty of obstacle trajectory distribution. The higher the entropy value, the stronger the constraint, forcing path points to converge towards the low-entropy region. The group cooperative collision avoidance behavior function models the repulsive force between multiple agents as a potential field function. The intensity of the repulsive force is jointly determined by the relative distance and relative velocity. It is recommended that the relative distance threshold range be 0.5 to 3.0 m. The specific value is determined by repeatedly conducting obstacle avoidance success rate experiments in obstacle scenarios with different densities and performing piecewise regression analysis on the experimental data. The recommended number of belief propagation iterations is 3 to 8 times, which is determined through multiple rounds of comparative experiments on convergence speed and computation delay. The steps for establishing the training dataset for the Conditional Random Field (CRF) deep extended spatiotemporal graph evolution model specifically include: collecting 3D scene simulation data containing randomly moving obstacles of different densities, recording the rover's pose sequence, the spatiotemporal trajectory of the obstacles, and the corresponding collision-free reference path, constructing a triplet dataset with a recommended data volume of no less than 50,000 trajectory samples; for any insufficient data, data augmentation is performed by applying random perturbations to existing trajectories; the original trajectory data undergoes graph structuring preprocessing under the topological graph sparse adjacency propagation mechanism to generate node feature matrices and adjacency matrix pairs. The steps for training the CRF deep extended spatiotemporal graph evolution model specifically include: using the composite energy function as the base term of the loss function, combined with a path smoothness regularization term, and employing an asynchronous parallel gradient accumulation strategy to update the network parameters of each subgraph channel; the recommended training epochs range from 80 to 200 epochs, and the recommended initial learning rate range is [missing value]. ~ The specific values ​​were determined through a learning rate warm-up experiment and an inflection point analysis of the energy function convergence curve on the validation set; early termination was triggered when the energy function on the validation set did not decrease for 10 consecutive rounds.

[0046] Among them, the linearized axial separation attention model for group collaborative collision avoidance behavior is based on deep learning. Its specific structure is as follows: The input layer receives the dynamic obstacle trajectory sequence tensor and the current state features of the rover. It projects the multi-source heterogeneous inputs to a unified, extremely low-dimensional latent space through a shared latent space low-dimensional alignment operator. The underlying core employs a linear attention approximation mechanism and an axial separation attention mechanism, decoupling the traditional fully connected attention matrix along... axis, axis, The system employs three uniaxial one-dimensional attention operators, each introducing a local sliding window with a suggested window size of 8–32 feature units, thus avoiding the quadratic computational cost of explicitly generating the complete association matrix. The behavioral repulsion function of the group's collaborative collision avoidance behavior is embedded as a weight scaling factor in the linear approximation operator, allowing the attention weights to dynamically scale with the obstacle's repulsion strength. Information flow is controlled in real-time by cascaded gating attention; lightweight cascaded gating intercepts and blocks worthless long-distance non-collision feature flows based on the repulsion field strength. A long-span residual dense jump matrix spans multiple axial attention layers, ensuring lossless transmission of the microscopic interactive behavior features of dynamic obstacles between layers. The output layer generates a smooth 3D roaming path point set tensor with rigid obstacle avoidance properties. The attention weight allocation between neurons, the activation intensity of residual jump paths between layers, the gating route update frequency of data loop, the batch size of each subgraph channel in memory allocation, and the parallel scheduling strategy of axial operators in hierarchical allocation are all jointly determined by the following sparse activation route dynamic adjustment function and particle swarm optimization sparse activation route adaptive update algorithm.

[0047] The steps for establishing the training dataset for the linearized axial separation attention model of the group cooperative collision avoidance behavior specifically include: generating a roaming scene containing dynamic obstacles of different densities and motion patterns in a 3D simulation environment; recording the obstacle trajectory tensor, roamer state features, and corresponding reference smooth path point sets; constructing a labeled dataset with a recommended data volume of no less than 80,000 samples; and augmenting the obstacle trajectories by applying Gaussian perturbations and random scaling to any insufficient data. The steps for training the linearized axial separation attention model of the group cooperative collision avoidance behavior specifically include: using the weighted sum of the mean square error of the path point set and the violation penalty term of the behavioral repulsion force as the loss function; the recommended weighting coefficient is within the range of 0.3 to 0.7, and the specific value is determined through multiple rounds of grid search experiments on the obstacle avoidance success rate of the validation set under different weighting coefficients; and adopting a hierarchical learning rate strategy, with the recommended learning rate range for the axial attention layer being [missing value]. ~ The recommended learning rate range for the gated routing layer is: ~ The recommended number of training rounds is 100 to 300 rounds. The optimal number of stopping rounds is determined by analyzing the joint convergence curve of the path smoothness index and obstacle avoidance success rate on the validation set.

[0048] The particle swarm optimization sparse activation route adaptive update algorithm is organically nested within the gated routing mechanism of the linearized axial separation attention model of swarm collaborative collision avoidance behavior. During particle swarm initialization, the position of each particle represents a sampling point in the gated routing weight parameter space, and the velocity represents the parameter update step size. The algorithm calculates the real-time prediction error rate under the current route output when inputting streaming features, and combines the prediction error rate with the path smoothness index to form a fitness function. In each iteration, each particle instantaneously adjusts its update step size based on its own historical best fitness and the swarm's global best fitness. The updated particle position is directly mapped to the real-time selection weight of the gated route. The suggested particle number range is 10–50, and the suggested number of iterations per round is 5–20. Specific values ​​are determined through multi-round comparative experiments on route convergence speed and computational latency. This algorithm frees the adjustment of network weights from the lag of solely relying on backpropagation, achieving millisecond-level dynamic adaptive route adjustment oriented towards scene changes.

[0049] The sparse activation routing dynamic adjustment function is used to adjust the sparse activation threshold parameter of the gated routing layer in the linearized axial separation attention model of group cooperative collision avoidance behavior; this function is based on the mean of the repulsive force field in the current frame. Path smoothness index Compared with the current prediction error rate Calculate the comprehensive routing pressure index using three data points. The calculation method is as follows ;in , , The weighting coefficients and The suggested range for all three coefficients is 0.2–0.6, and the specific values ​​were determined through multiple rounds of experiments and variance analysis on the accuracy of routing decisions under different weight combinations; when Different adjustment strategies are used for the sparse activation threshold parameter of the gated routing layer when the ranges are different: when When the sparse activation threshold is raised to the high sparsity range, it forcibly blocks the flow of low-repulsion long-distance features, reducing the computational cost of attention weight allocation between neurons; when At this time, the sparse activation threshold remains in the moderately sparse range, and the activation intensity of the residual jump path between layers is based on the current... Value linear interpolation adjustment; when At that time, the sparse activation threshold is reduced to the low sparsity range, all axial attention operators are activated and the local sliding window size is expanded. At the same time, the gating route update frequency of the data loop is increased to the highest level, and the batch size of each subgraph channel in the memory allocation is compressed to ensure real-time performance.

[0050] Among them, the long-span residual dense jump matrix refers to the dense cross-layer connection weight matrix established between multiple non-adjacent layers of the network. Each cross-layer connection path carries an independent learnable scaling coefficient, enabling deep networks to directly acquire high-resolution spatial features from shallow layers. The adaptive exponential decay sliding window is a sliding time window mechanism that retains historical key viewpoint features by weighting them according to an exponential decay coefficient. The weight of historical features further away from the current frame is lower. The suggested range of the decay coefficient is 0.85 to 0.99. The specific value is determined through multiple rounds of experiments on the topological coherence of long-distance paths under different decay coefficients and statistical analysis of trajectory continuity indicators. Global topological coherence refers to the global geometric continuity of the roaming path without breaks or detours under long-term and large-spatial spans.

[0051] Among them, asynchronous lightweight local subgraph refers to the independent subgraph unit obtained by dividing the three-dimensional spatial topology into spatial partitions. Each subgraph only maintains the nodes and edges within its own partition. The boundary features of each subgraph are synchronized through an asynchronous message passing mechanism, without needing to access the global shared data structure. Asynchronous parallel pipeline multi-channel refers to a parallel scheduling structure in which multiple subgraph inference channels are executed in an interleaved manner without waiting for each other to complete. Each channel independently holds a local buffer, eliminating thread mutual exclusion waiting caused by global data locks.

[0052] Among them, the Conditional Random Field Deep Extension Spatiotemporal Graph Evolution Model extends the single-layer potential function of the traditional graph-theoretic Markov Random Field into a multi-level deep cascaded potential function, enabling the model to refine the spatiotemporal interaction probability between obstacles and roamers layer by layer on the graph structure. The Local Asymmetric Gated Recurrent Unit refers to an improved gated recurrent unit in which the input gate and forget gate weight matrices are not shared. Through asymmetric weight design, the unit has different selectivity for input features and historical states, which improves the ability to capture short-term changes in obstacle trajectories. The model eliminates multi-threaded global lock waiting through an asynchronous parallel pipeline multi-channel structure. Each channel's belief propagation iteration converges independently before lightweight boundary alignment is performed, so that the convergence of the global energy function can be completed in a very short time after removing thread waiting losses. At the same time, the cascaded gating early interception mechanism greatly reduces the inference overhead of invalid feature branches, making the overall computational resource consumption significantly lower than the traditional full-graph synchronous inference method. The final output roaming path has better obstacle avoidance smoothness and global rationality in a multi-dynamic obstacle environment.

[0053] Among them, the shared latent space low-dimensional alignment operator refers to the alignment operation operator that applies independent linear projections to multi-source heterogeneous inputs and then concatenates the projection results into the same low-dimensional latent space, so that different modal features have a unified dimensional benchmark in subsequent attention calculations; the linear attention approximation mechanism refers to the approximation calculation method that reduces the quadratic computational complexity of standard attention to linear level by using kernel function decomposition. By expanding the kernel function of query and key, calculating the key-value product first and then multiplying it with the query, the overhead of explicitly constructing the complete attention matrix is ​​avoided; the rigid obstacle avoidance attribute refers to the attribute that the path point set satisfies the constraint of maintaining a safe distance from obstacles at any time. The recommended range of safe distance is 0.3 to 2.0 m, and the specific value is determined by conducting collision rate statistical experiments under different combinations of obstacle size and movement speed.

[0054] In an optional implementation, the above-mentioned method for dynamically generating a 3D scene roaming path can also be implemented by a computer to form a 3D scene roaming path dynamic generation system. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute all the steps of the above method. The computer is connected to a data acquisition device for providing streaming 3D mesh and point cloud data and a display output device for carrying out roaming interactive operations.

[0055] The specific implementation of step S01 is as follows: At the streaming data inlet, the input 3D mesh data and point cloud data undergo pre-orthogonal decoupling processing. Specifically, an orthogonal basis decomposition algorithm is used to perform orthogonal decomposition on the high-dimensional mesh feature matrix, converting linearly correlated feature components between dimensions into mutually orthogonal independent components, eliminating the interference of redundant information on subsequent mapping. After orthogonal decoupling, a dynamic lightweight topology structure is constructed. This graph structure adjusts the number of nodes, edges, and connections in real time according to the spatial distribution of the input data, automatically reducing the topology scale to reduce computational overhead in high-load scenarios. Subsequently, the high-dimensional features after orthogonal decoupling are compressed to a low-dimensional latent space of 2% to 8% of the original feature dimension through local nonlinear state space mapping. The specific low-dimensional ratio is determined by jointly measuring reconstruction error and computational latency on multiple sets of 3D scene data of different scales and performing Pareto optimality analysis on the experimental data. Finally, the multi-source heterogeneous input data completes topological alignment in the low-dimensional latent space, providing a feature representation with a unified dimensional benchmark for subsequent steps.

[0056] The specific implementation of step S02 is as follows: The rendering features within the current view frustum are extracted in real time and used as input to the forward channel of path prediction. A learnable mask dynamic sparse perception mechanism configures a set of binarized mask weights for each feature channel of the neural network, which can be updated during training. During the inference phase, the corresponding channels are dynamically activated or masked based on the real-time changes in the view frustum's geometric parameters, ensuring that the network only performs calculations on core features strongly correlated with the current view frustum, masking redundant feature flows unrelated to the current frame. Local activation operators perform nonlinear activation operations only within the local spatial neighborhood of the feature map, avoiding redundant computational overhead caused by global activation. When a sudden change occurs in the rendering features within the view frustum, the forward channel of path prediction triggers an instantaneous adaptive jump mechanism. The forward propagation process directly skips network layers unrelated to the current features, routing the activation results to subsequent output layers, thus achieving dynamic routing adjustment of the forward channel.

[0057] The specific implementation of step S03 is as follows: The current pose sequence of the 3D rover and the spatiotemporal trajectory sequence of dynamic obstacles are taken as inputs. Through a topological graph sparse adjacency transmission mechanism, these are represented as nodes and directed edges of the graph, respectively, constructing a full graph representation and inputting it into a conditional random field deep-extended spatiotemporal graph evolution model. This model extends the single-layer potential function of the traditional Markov random field to a multi-level deep cascaded potential function, which can refine the spatiotemporal interaction probability between obstacles and the rover layer by layer. The energy function is composed of a state entropy-increasing dynamic constraint function and a group cooperative collision avoidance behavior function. The state entropy-increasing dynamic constraint function quantifies the uncertainty of the obstacle trajectory distribution; the higher the entropy value, the stronger the constraint, forcing path points to converge towards low-entropy regions. The group cooperative collision avoidance behavior function models the repulsive force of multi-agent behavior as a potential field function. The intensity of the repulsive force is jointly determined by the relative distance and relative velocity, with a relative distance threshold range of 0.5–3.0 m. The information flow is conditionally controlled by cascaded gating of the potential function; ultra-lightweight cascaded gating intercepts low-potential feature branches in advance. Each sub-layer independently performs belief propagation iterations through an asynchronous parallel pipeline with multiple channels, ranging from 3 to 8 iterations. Each channel utilizes locally asymmetric gated recurrent units, where the input and forget gate weight matrices are not shared, enhancing the ability to capture short-term abrupt changes in obstacle trajectories. After deep cascading correction of the multi-layer potential function, the global energy function converges, outputting a multi-track obstacle avoidance roaming curve with rigid obstacle avoidance properties.

[0058] The specific implementation of step S04 is as follows: The dynamic obstacle trajectory sequence tensor and the current state features of the rover are synchronously input into the linearized axially separated attention model of the group cooperative collision avoidance behavior. The input layer applies independent linear projections to the multi-source heterogeneous inputs through a shared latent space low-dimensional alignment operator, and then stitches the projection results into a unified low-dimensional latent space. The core of the bottom layer adopts a linear attention approximation mechanism, which reduces the quadratic computational complexity of standard attention to linear level through kernel function decomposition, and decouples the attention matrix into one-dimensional attention calculation operators along three uniaxial axes: X-axis, Y-axis, and Z-axis. Each operator introduces a local sliding window with a window size ranging from 8 to 32 feature units. The behavioral repulsion force function of the group cooperative collision avoidance behavior is embedded in the linear approximation operator in the form of a weight scaling factor, so that the attention weights are dynamically scaled with the obstacle repulsion force intensity. The particle swarm optimization sparse activation route adaptive update algorithm uses a fitness function composed of a real-time prediction error rate and a path smoothness index. In each iteration, each particle adjusts its update step size based on its own historical best fitness and the global best fitness of the swarm. The number of particles ranges from 10 to 50, and the number of iterations per round ranges from 5 to 20. The updated particle positions are directly mapped to the gated route real-time selection weights. The output layer generates a smooth 3D roaming path point set tensor. The path point set satisfies the rigid obstacle avoidance property with a safe distance ranging from 0.3 to 2.0 m.

[0059] The specific implementation of step S05 is as follows: A long-span residual dense jump matrix establishes a dense cross-layer connection weight matrix between multiple non-adjacent layers of the network. Each cross-layer connection path carries an independent learnable scaling coefficient, enabling the deep network to directly acquire high-resolution spatial features from shallow layers and losslessly transmit historical key viewpoint features across layers. An adaptive exponential decay sliding window retains historical key viewpoint features weighted by an exponential decay coefficient. The weight of historical features farther from the current frame is lower, with the decay coefficient ranging from 0.85 to 0.99. The specific value was determined through multiple rounds of experiments on the topological coherence of long-distance paths under different decay coefficients and statistical analysis of trajectory continuity indicators. The synergistic effect of these two mechanisms enables the roaming path to maintain unbroken and detour-free global geometric continuity over long and large spatial spans, achieving global topological coherence.

[0060] The specific implementation of step S06 is as follows: The three-dimensional spatial topology is divided into asynchronous lightweight local subgraphs according to spatial partitions. Each subgraph only maintains the nodes and edges within its own partition. Boundary features are synchronized between subgraphs through an asynchronous message passing mechanism, without needing to access the globally shared data structure. The inference channels of each subgraph are executed interleaved under an asynchronous parallel pipeline multi-channel scheduling structure. Each channel independently holds its local buffer, eliminating thread mutual exclusion waiting caused by global data locks. After each subgraph completes independent inference, lightweight boundary alignment is performed, and the output results of each subgraph are aggregated to generate a globally consistent final roaming path.

[0061] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for dynamically generating a roaming path in a three-dimensional scene.

[0062] A third aspect of the present invention provides a dynamic roaming path generation system in a three-dimensional scene, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

[0063] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that it synergistically cuts off the computational coupling bottleneck between high-dimensional multi-source heterogeneous data fusion and global path inference from four dimensions: data dimensionality reduction, feature sparsification, graph inference parallelization, and attention linearization. First, the pre-orthogonal decoupling removes the linear correlation between the dimensions of high-dimensional grid features through orthogonal basis decomposition, making the feature components independent of each other. The dynamically lightweight topology then maps the independent components to a low-dimensional latent space of 2% to 8% of the original dimension. All subsequent calculations are performed in this extremely low-dimensional space, fundamentally compressing the computational workload of multi-source data fusion to a level that can be processed in real time. Second, the learnable mask dynamic sparse perception mechanism configures binary mask weights for each feature channel. During inference, the corresponding channels are activated or masked in real time according to the geometric changes of the view frustum. Combined with the local activation operator, nonlinear activation is performed only in the local spatial neighborhood, so that the forward channel can skip irrelevant network layers and directly route to the output layer when the view frustum changes abruptly, thereby significantly reducing the inference latency per frame without sacrificing the path prediction accuracy. Furthermore, the Conditional Random Field Deep Extension Spatiotemporal Graph Evolution Model represents the rover's pose and the spatiotemporal trajectory of dynamic obstacles as graph nodes and directed edges. Graph inference is driven by a composite energy function consisting of a state entropy increase dynamic constraint function and a group cooperative collision avoidance behavior function. The multi-level potential function deep cascade correction enables the model to refine the spatiotemporal interaction probability between obstacles and the rover layer by layer. The asynchronous parallel pipeline multi-channel structure allows each subgraph to independently perform belief propagation iterations, eliminating thread mutual exclusion waiting caused by global data locks. After each channel converges, only lightweight boundary alignment is needed to output a globally consistent obstacle avoidance path, so that the convergence time of the global energy function in multi-dynamic obstacle scenarios is not affected by thread waiting overhead. Finally, the linearized axial separation attention model for collaborative collision avoidance decouples the traditional fully connected attention matrix into one-dimensional linear attention operators along three uniaxial axes: X, Y, and Z. Combined with a local sliding window, this avoids the quadratic overhead of explicitly generating a complete association matrix. The particle swarm optimization sparse activation route adaptive update algorithm dynamically adjusts the gated route weights at the millisecond level, freeing network weight updates from the lag of relying solely on backpropagation and enabling real-time route adjustment in response to scene changes. The long-span residual dense jump matrix and adaptive exponential decay sliding window further ensure the global topological coherence of long-range paths, maintaining the global geometric continuity of the generated roaming path without breaks or detours across large spatial spans. The synergistic effect of these mechanisms ensures that the entire process from data input to path output is completed under low computational overhead constraints, logically proving that the present invention can solve the core technical problem of real-time fusion of multi-source heterogeneous data and simultaneous global path planning in environments with multiple dynamic obstacles.

[0064] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0065] The specific implementation method of step S01 is as follows.

[0066] For the incoming grid and point cloud data, firstly, at the boundary of the streaming data stream, the high-dimensional grid feature matrix is... Perform orthogonal basis decomposition, where The number of feature samples, The original feature dimension. Singular value decomposition is then used to... Decomposed into , and These are left and right orthogonal matrices, As a diagonal singular value matrix, after removing the linear correlation between dimensions, each feature component is independent, avoiding redundant calculations in subsequent mappings. The dynamic lightweight topology structure is constructed based on the principle of spatial self-organization evolution. , where the node set With edge set The size of the feature vector is adjusted in real time according to the spatial distribution of the input data. Subsequently, the orthogonally decoupled high-dimensional features are compressed into a low-dimensional latent space through local nonlinear state-space mapping. The compressed low-dimensional feature vector... The mapping relationship formula is expressed as follows:

[0067] ;

[0068] In the formula, The characteristic matrix after orthogonal decoupling. Here is the encoder weight matrix. For bias vectors, It is a local nonlinear activation function. To reconstruct the error term, The dimension of the low-dimensional latent space, whose value range is equal to the original feature dimension. The percentage of reconstruction error to computational delay was determined through a joint Pareto-optimal experiment conducted on multiple sets of 3D scene data of varying scales, ranging from 2% to 8%. The left side of the equation... With the right side All outputs are Dimensional vectors with consistent units. After topological alignment, heterogeneous data from multiple sources are unified in a low-dimensional latent space. This indicates that it is for use in subsequent steps.

[0069] The specific implementation method of step S02 is as follows.

[0070] Rendering feature map within the current view frustum ,in For the number of channels, and These represent the spatial height and width of the feature map, respectively, and can be obtained through a learnable mask vector. Dynamic sparse perception retains only core features. Mask weights are updated during training, and during inference, corresponding channels are activated or masked in real-time based on changes in the view frustum geometry. The sparsed features are... ,in To broadcast multiplication channel by channel, This is the sparsified feature map. Local activation operators only apply to the local spatial neighborhood. Non-linear activation is performed within the frustum to avoid redundant computational overhead from global activation. When a sudden change occurs in the rendered features within the frustum, the forward propagation process skips network layers unrelated to the current features and directly routes the activation results to the subsequent output layer, achieving instantaneous adaptive switching of the forward path prediction channel and improving inference real-time performance.

[0071] The specific implementation method of step S03 is as follows.

[0072] The current pose sequence of the 3D rover With dynamic obstacle spatiotemporal trajectory sequence tensor Represented as a graph The nodes and directed edges, where For the first The pose vector of the time-space roamer. for The obstacle is in the first... The spacetime state tensor at time t, For timing length, For the spatiotemporal graph node set, It is a set of directed edges. An adjacency matrix is ​​constructed using the sparse adjacency propagation mechanism of the topological graph. The formula for information transmission in a graph is expressed as follows:

[0073] ;

[0074] In the formula, For the first Layer nodes The hidden state vector. For the hidden state dimension, For the first Layer learnable weight matrix, For the first Layer bias vector, For nodes The set of neighboring nodes, For nodes in the adjacency matrix With nodes Connection weights between them and They are nodes With nodes The degree, It is a nonlinear activation function. The composite energy function is determined by the state entropy increasing dynamic constraint function. Group cooperative collision avoidance behavior function With path smoothness regularization term It is a composite material, and the formula is expressed as follows:

[0075] ;

[0076] In the formula, , , These are the weighting coefficients for each item; all three are dimensionless positive real numbers. , , , The dimensions of each quantity are unified into a dimensionless energy scalar through internal normalization. The state entropy increasing dynamic constraint function quantifies the uncertainty of the obstacle trajectory distribution, and the formula is as follows:

[0077] ;

[0078] In the formula, For the first The obstacle is in the first... The probability distribution value of the location at any given time is represented by a dimensionless information entropy function. Higher entropy values ​​indicate stronger constraints, forcing pathpoints to converge towards lower entropy regions. The group cooperative collision avoidance behavior function models the repulsive force of multi-agent behavior as a potential field function, expressed by the following formula:

[0079] ;

[0080] In the formula, For the wanderer in the Time and location With the Obstacle locations The Euclidean distance between them, in meters. For the Wanderer and the First The obstacle is in the first... The relative velocity amplitude at time t, in units of , The repulsive potential field function is whose output is normalized to a dimensionless scalar, and the repulsive force strength is given by... and The relative distance threshold is determined jointly to be within the range of 0.5–3.0 m. (Input gate of the locally asymmetric gated loop unit) With the Gate of Oblivion The weight matrix is ​​not shared, and the formula is expressed as follows:

[0081] ;

[0082] ;

[0083] In the formula, , Here are the input weight matrices for the input gate and the forget gate. , This is the corresponding hidden state weight matrix. Input the feature vector at the current time. For the input feature dimension, This is the hidden state vector of the gated recurrent unit at the previous time step. For the hidden state dimension of the gated loop unit, , The input and forget gate bias vectors are used, and the asymmetric weight design allows units to have different selectivity for input features and historical states, improving the ability to capture short-term changes in obstacle trajectories. The model's training dataset contains no fewer than 50,000 trajectory sample triples, with 80–200 training epochs and an initial learning rate of [missing information]. ~ The verification set energy function is not reduced for 10 consecutive rounds, triggering an early stop.

[0084] The specific implementation method of step S04 is as follows.

[0085] The low-dimensional alignment operator in the shared latent space will align the dynamic obstacle trajectory sequence tensor. Characteristics of the current state of the wanderer After being independently linearly projected, they are stitched together into a unified low-dimensional latent space, as expressed by the following formula:

[0086] ;

[0087] In the formula, and For each modal projection matrix, and For the corresponding bias vector Let the obstacle tensor be flattened vector. For the flattened dimensions, For the roamer's state feature dimension, The dimensions are the projected dimensions of each modality. The aligned low-dimensional feature vectors This is a vector concatenation operation. The axial separation attention mechanism decouples the traditional fully connected attention into an axial attention operation. axis, axis, The one-dimensional attention calculation along the three axes is expressed by the following formulas for attention along each axis:

[0088] ;

[0089] In the formula, , , These are the query, key, and value matrices along a certain axis. The sequence length along this axis. The dimension of the key vector. The value vector dimension is used, and the local sliding window size ranges from 8 to 32 feature units. The linear attention approximation mechanism utilizes kernel function decomposition to reduce the quadratic computational complexity to linear time, avoiding the overhead of explicitly constructing the complete attention matrix by first calculating the key-value product and then multiplying it with the query. Behavioral repulsion potential field function. Embedded as a weight scaling factor in the linear approximation operator, the attention weights are dynamically scaled according to the strength of the obstacle repulsion force. The fitness function formula for the particle swarm optimization sparse activation route adaptive update algorithm is expressed as follows:

[0090] ;

[0091] In the formula, For the first The fitness value of each particle. For the first The real-time prediction error rate for each particle under the routing weight. For the corresponding path smoothness index, and The weighting coefficients and All three are dimensionless scalars with consistent dimensions. The formulas for updating particle velocity and position are expressed as follows:

[0092] ;

[0093] ;

[0094] In the formula, and The first The particle in the first The velocity vector and position vector of each iteration both belong to the gated routing weight parameter space. The inertia weight has an empirical value of 0.7. and This is the learning factor, typically set to 2.0. and for Uniformly distributed random numbers in an interval For the first The best historical position of each particle The global optimal position for all particles is determined, with the number of particles ranging from 10 to 50, and the number of iterations per round ranging from 5 to 20. The training loss function is a weighted sum of the mean squared error of the path point set and the violation penalty term for behavioral repulsion, with weighting coefficients ranging from 0.3 to 0.7. The learning rate of the axial attention layer ranges from [missing value]. ~ The learning rate range of the gated routing layer is ~ The number of training rounds is 100 to 300, and the number of training samples is no less than 80,000.

[0095] The specific implementation method of step S05 is as follows.

[0096] The sparse activation routing dynamic adjustment function is based on the mean value of the repulsive force field in the current frame. Path smoothness index Compared with the current prediction error rate Calculate the comprehensive routing pressure index The formula is expressed as follows:

[0097] ;

[0098] In the formula, , , The weighting coefficients and The recommended range for all three coefficients is 0.2 to 0.6. Normalized to interval, , Thus As a dimensionless scalar, the values ​​of each coefficient were determined through multiple rounds of experiments and variance analysis on the routing decision accuracy under different weight combinations. When the sparse activation threshold is increased to the high sparsity range, it forcibly blocks the flow of long-distance features with low repulsion; when At that time, the sparse activation threshold remains in the moderately sparse range, and the activation intensity of the interlayer residual jump path is adjusted according to... Value linear interpolation adjustment; when At this time, the sparse activation threshold is reduced to the low-sparse range, all axial attention operators are activated and the local sliding window size is expanded. Simultaneously, the gated route update frequency is increased to the highest level, and the batch size of each subgraph channel is compressed to ensure real-time performance. Long-span residual dense jump matrices exist in multiple non-adjacent layers. and To establish cross-layer connections, the formula is as follows:

[0099] ;

[0100] In the formula, For the first The output hidden state vector of the layer. For the first The original output hidden state vector of the layer. This is the output hidden state vector after cross-layer connection. For the first Layer to the first The learnable scaling factor for inter-layer connectivity pathways is a dimensionless scalar, and both sides of the equation are equal to 1. A dimensionless vector with consistent dimensions. An adaptive exponentially decaying sliding window retains historical key viewpoint features weighted by an exponential decay coefficient, as expressed in the following formula:

[0101] ;

[0102] In the formula, The current moment is a weighted historical viewpoint feature vector. The distance from the current frame The historical key viewpoint feature vector of the step, The length of the sliding window. The exponential decay coefficient, ranging from 0.85 to 0.99, was determined through multiple rounds of experiments on the topological coherence of long-distance paths under different decay coefficients and statistical analysis of trajectory continuity indices. Historical feature weights are determined by distance from the current frame. The lower the value, the better to maintain the global topological coherence of long-distance paths.

[0103] The specific implementation method of step S06 is as follows.

[0104] The three-dimensional spatial topology is divided into spatial partitions. An asynchronous lightweight local subgraph Each subgraph maintains only the nodes and edges within its own partition. The inference channels of each subgraph execute interleaved in an asynchronous parallel pipeline without waiting for each other to complete. Each channel independently holds its local buffer, eliminating thread mutual exclusion caused by global data locks. After each subgraph's belief propagation iteration converges independently, lightweight boundary alignment is performed. The number of belief propagation iterations ranges from 3 to 8, determined through multiple rounds of comparative experiments comparing convergence speed and computational latency. Finally, the outputs of each subgraph are merged to generate the final roaming path. Path points must maintain a safe distance threshold from obstacles at any given time. The safe distance ranges from 0.3 to 2.0 meters, determined through statistical experiments on collision rates under different combinations of obstacle sizes and movement speeds.

[0105] To better understand and implement this invention, a specific application scenario of the invention is provided in Embodiment 2 below: This embodiment selects a large-scale 3D virtual scene containing multi-density dynamic obstacles as the test environment. The scene space ranges from 500m to 500m to 50m, and the real-time access rate of the mixed 3D mesh data and point cloud data is approximately 120 frames per second. A total of 80 dynamic obstacles are set within the scene, with motion modes covering three categories: uniform linear motion, random polygonal motion, and group aggregation motion. The straight-line distance between the start and end points of the roamer is approximately 350m. The objective of this embodiment is to verify whether the method of the present invention can generate a smooth roaming path that satisfies rigid obstacle avoidance properties in real time within this scene.

[0106] In step S01, the input grid data has a single-frame feature dimension of 4096 dimensions, and the point cloud data has a single-frame feature dimension of 2048 dimensions. After performing pre-orthogonal decoupling on both data streams, the linear correlation coefficients between each dimension are reduced to below 0.03, and the feature components tend to be independent of each other. The dynamic lightweight topology automatically adjusts the number of nodes according to the current frame spatial distribution. In this embodiment, the average number of nodes is reduced from the initial 8200 to 1450. Subsequently, through local nonlinear state space mapping, the grid features are compressed to 82 dimensions (approximately 2% of the original dimensions), and the point cloud features are compressed to 82 dimensions. The two features are then topologically aligned within the same low-dimensional latent space, as shown below. Figure 2 The diagram shows the data flow of pre-orthogonal decoupling and dimensionality reduction alignment.

[0107] In step S02, the learnable mask dynamic sparse perception mechanism configures binarized mask weights for each of the 82 channels of the mesh feature. In the frame sequence of this embodiment, the view frustum coverage is approximately 30% of the entire scene, the masking mechanism activates an average of 24 channels, and the masking ratio is approximately 71%. When the view frustum deflects rapidly, the path prediction forward channel triggers an instantaneous adaptive jump, skipping an average of 3 network layers, significantly reducing the forward computation latency compared to full-channel activation. The local activation operator performs nonlinear activation within a 3×3 spatial neighborhood, avoiding redundant computations of global activation.

[0108] In step S03, the spatiotemporal trajectory sequence tensors of 80 dynamic obstacles and the pose sequence of the rover jointly construct a graph structure. The graph contains 81 nodes and approximately 620 directed edges. The network structure of the conditional random field depth-extended spatiotemporal graph evolution model in this embodiment is as follows: Figure 3As shown, the model consists of six multi-level deep cascaded potential function layers, with the belief propagation iteration count set to five. The average entropy of the obstacle trajectory distribution in the current frame is calculated using the state entropy increase dynamic constraint function, resulting in an average entropy of 1.82. The relative distance threshold in the group cooperative collision avoidance behavior function is set to 1.5m. This embodiment uses four asynchronous parallel pipeline channels. Each channel independently completes the belief propagation iteration and then performs lightweight boundary alignment, outputting 12 candidate multi-track obstacle avoidance roaming curves. Table 1 shows the belief propagation convergence rounds and the corresponding number of candidate paths for each channel.

[0109] Table 1. Convergence of Belief Propagation and Statistical Analysis of Candidate Paths in Each Flow Channel

[0110]

[0111] In step S04, the linearized axial separation attention model for collaborative collision avoidance receives the dynamic obstacle trajectory sequence tensor and the current state features of the rover as input. The shared latent space low-dimensional alignment operator projects the two inputs into a unified 82-dimensional latent space. The axial separation attention mechanism decomposes the attention computation into three one-dimensional operators along the X, Y, and Z axes, setting the local sliding window size to 16 feature units, reducing the computational complexity from quadratic to linear. The particle swarm optimization sparse activation route adaptive update algorithm initializes the number of particles to 30 and iterates 10 times per round. In the sparse activation route dynamic adjustment function, the weight coefficients... , , The overall routing pressure index in frame 50 of this embodiment. The threshold is within the medium pressure range, the sparse activation threshold remains in the medium sparse range, and the proportion of activated neurons in the gated routing layer is approximately 58%. Figure 4 The diagram shows the generation process of the path point set tensor. The output smooth 3D roaming path point set tensor contains 1280 path points. The minimum distance between the path points and obstacles is no less than 0.8m, which satisfies the rigid obstacle avoidance property. Table 2 shows the statistics of the safe distance of path points in different obstacle density sections.

[0112] Table 2. Statistics on safe distances to path points in sections with different obstacle densities.

[0113]

[0114] In step S05, the long-span residual dense jump matrix establishes cross-layer connections between layers 2 and 8, and layers 3 and 10 of the model, respectively, with each path carrying an independent learnable scaling coefficient. The decay coefficient of the adaptive exponential decay sliding window is set to 0.92, the window covers 30 historical frames, and the historical feature weights from the 30th frame since the current frame decay to approximately 0.07. Figure 5The diagram illustrates the characteristic propagation structure of a long-span residual dense jump matrix and an adaptive exponential decay sliding window. In the 350m full-length path, the path topology continuity index (number of trajectory breakpoints) is 0, verifying the ability to maintain global topology coherence. As the weight of historical frames decreases exponentially with the decay coefficient, the influence of distant historical key viewpoints on the current path gradually diminishes, and the path smoothness curve tends to stabilize, indicating that this mechanism effectively suppresses the interference of historical noise on long-range path planning.

[0115] In step S06, the 500m×500m×50m three-dimensional space is divided into 25 asynchronous lightweight local subgraphs, each covering a 100m×100m×50m spatial partition. This embodiment deploys four asynchronous parallel pipeline channels, each independently holding a local buffer. After inference of each subgraph is completed, boundary features are synchronized through an asynchronous message passing mechanism. For example... Figure 6 The diagram shows the inference scheduling structure of an asynchronous parallel pipeline with multiple channels. There are no global data locks waiting between channels, and the thread waiting time during the inference process is 0ms. Finally, a globally consistent roaming path is output. Table 3 shows the inference latency statistics for each subgraph.

[0116] Table 3. Statistics on inference delay for each subgraph

[0117]

[0118] The network structures of the two artificial intelligence models in this invention are further explained in the embodiments as follows. The main network structure of the Conditional Random Field Deep Extended Spatiotemporal Graph Evolution Model consists of a graph embedding layer, six deep cascaded potential function layers (each layer containing a topological graph sparse adjacency propagation module and a locally asymmetric gated loop unit), an asynchronous parallel pipeline multi-channel scheduling module, and an output layer. The data flow is as follows: the input node feature matrix and adjacency matrix are converted into node embedding vectors by the graph embedding layer, and then energy calculation and belief propagation are performed layer by layer through the deep cascaded potential function layers. After independent iteration of each channel, candidate path tensors are generated by the output layer. The main structure of the group cooperative collision avoidance linearized axial separation attention model consists of a shared latent space low-dimensional alignment operator layer, three axial one-dimensional linear attention operator layers (each containing a local sliding window), a long-span residual dense jump matrix connection layer, a cascaded gated attention control layer, and a path point output layer. The data flow is as follows: the two inputs are projected to 82 dimensions by the alignment operator, and then pass through the X-axis, Y-axis and Z-axis attention operator layers in sequence. The residual dense jump matrix transmits shallow features across layers. The gated routing layer dynamically adjusts the activation path under the drive of the particle swarm optimization sparse activation routing adaptive update algorithm, and finally outputs the path point set tensor.

[0119] For alternative network structures, two equivalent alternatives are given below. The first alternative is to replace the axial separation attention mechanism with a graph attention network structure. Based on graph node embedding, the multi-agent relationships in the obstacle trajectory sequence tensor are modeled. The data flow is as follows: the input features are projected through the low-dimensional alignment operator of the shared latent space, the neighborhood node information is aggregated through the multi-head graph attention layer, and then the path point set tensor is output through residual connections and gated routing layers. This alternative structure is consistent with the main scheme in terms of feature extraction dimension. The convergence objective of the loss function is the weighted sum of the mean square error of the path point set and the violation penalty term of the behavioral repulsion force. This is an equivalent replacement that can be implemented by those skilled in the art without creative effort. The second alternative is to replace the locally asymmetric gated recurrent units in the deep extended spatiotemporal graph evolution model of the conditional random field with a temporal convolutional network module. This module uses one-dimensional causal convolution to model the temporal dependencies of obstacle trajectory sequences. The data flow is as follows: node embedding vectors are sequentially passed through multiple causal convolutional layers to extract temporal features. The convolutional output and the graph embedding layer output are added together by residuals and then enter the deep cascaded potential function layer. The subsequent process is the same as the main solution. This alternative structure maintains the same dimensionality for processing input data as the main solution and is suitable for scenarios with high requirements for temporal modeling capabilities of trajectory sequences.

[0120] The final roaming path generated by this invention has zero topological breakpoints within a 350m range, and the safe distance between all path points and obstacles is no less than 0.8m. As the obstacle density increases from 0.12 per segment in section 1... Gradually increasing to 0.61 in section 4. The minimum safe distance between path points converged from 1.42m to 0.80m, and the safe distance curve tended to be stable rather than continuing to decrease. This indicates that the group cooperative collision avoidance behavior function in the composite energy function maintained an effective repulsive force constraint in the high-density obstacle area, keeping the path points above the safe distance threshold. From a technical principle perspective, this invention verifies the ability of the present invention to fuse multi-source heterogeneous data in real time and simultaneously achieve high-quality global path planning in a multi-dynamic obstacle environment.

[0121] It should be noted that the variables involved in this invention are explained in detail in Tables 4, 5, and 6.

[0122] Table 4. Variable Explanation Table (Part 1)

[0123]

[0124] Table 5. Variable Explanation Table (Part Two)

[0125]

[0126] Table 6. Variable Explanation Table (Part 3)

[0127]

[0128] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamically generating a roaming path in a 3D scene, characterized in that, Includes the following steps: Pre-orthogonal decoupling is performed on the mesh and point cloud data flowing into the 3D scene, and the high-dimensional mesh features are mapped to the low-dimensional latent space using a dynamic lightweight topology structure, thus completing the topology alignment and lossless dimensionality reduction of multi-source heterogeneous data. The rendering features within the current view frustum are extracted, and a learnable mask dynamic sparse perception mechanism is used to retain only the core features of the view frustum. Combined with dynamic reconstruction using local activation operators, instantaneous adaptive jumping of the forward channel of path prediction is achieved. The current pose of the 3D rover and the spatiotemporal trajectory sequence of dynamic obstacles are input into the conditional random field depth-extended spatiotemporal graph evolution model. The composite energy function driven by the state entropy increase dynamic constraint function and the group cooperative collision avoidance behavior function is used to infer the multi-track obstacle avoidance rover curve in a multi-channel asynchronous parallel pipeline. The dynamic obstacle trajectory sequence tensor and the current state features of the roamer are input into the linearized axial separation attention model of the group collaborative collision avoidance behavior. The gated route weights are adjusted in real time by the particle swarm optimization sparse activation route adaptive update algorithm, and the output is a smooth three-dimensional roaming path point set tensor. By using a long-span residual dense jump matrix to transmit historical key viewpoint features across layers, and combining it with an adaptive exponential decay sliding window to lock long-distance spatial features, the global topological coherence of long-distance paths is maintained. The three-dimensional spatial topology is decoupled into asynchronous lightweight local subgraphs under a multi-dimensional network transmission and fusion mechanism. Each subgraph thread independently infers through an asynchronous parallel pipeline with multiple channels, eliminating thread waiting caused by global data locks and outputting the final roaming path.

2. The method for dynamically generating roaming paths according to claim 1, characterized in that, The aforementioned pre-orthogonal decoupling refers to performing orthogonal basis decomposition on the high-dimensional grid features at the boundary of the streaming data stream, removing the linear correlation between dimensions, making the feature components independent of each other, thereby avoiding redundant calculations in subsequent mapping.

3. The method for dynamically generating roaming paths according to claim 2, characterized in that, The dynamic lightweight topology refers to a graph structure constructed based on the principle of spatial self-organization and evolution. The number of nodes and edges and their connection relationships are adjusted in real time according to the spatial distribution of the input data, and the topology scale is automatically reduced under different load conditions. The dimensional range of the low-dimensional potential space is an extremely low-dimensional feature representation space within the low-dimensional ratio threshold range of the original feature dimensions. The low-dimensional ratio threshold is determined by conducting joint experiments on reconstruction error and computation delay on multiple sets of three-dimensional scene data of different scales and performing Pareto optimality analysis on the experimental data.

4. The method for dynamically generating roaming paths according to claim 3, characterized in that, The learnable mask dynamic sparse perception mechanism refers to configuring a set of binary mask weights that can be updated with training for each feature channel inside the neural network, and activating or blocking the corresponding channel in real time according to the geometric changes of the view frustum during inference. The local activation operator refers to an operator that performs nonlinear activation operations only within the local spatial neighborhood of the feature map; the instantaneous adaptive jump of the path prediction forward channel refers to the forward propagation process skipping network layers unrelated to the current features when a sudden change occurs in the rendered features within the view frustum, and directly routing the activation result to the subsequent output layer.

5. The method for dynamically generating roaming paths according to claim 4, characterized in that, The specific structure of the Conditional Random Field Deep Extended Spatiotemporal Graph Evolution Model is as follows: The input layer represents the 3D roamer pose sequence and the spatiotemporal trajectory of dynamic obstacles as nodes and directed edges of a graph, and constructs a full graph representation through a topological graph sparse adjacency transmission mechanism; the energy function is composed of a state entropy increase dynamic constraint function and a group cooperative collision avoidance behavior function, and is substituted into the multi-level potential function for calculation; the information flow direction within the network is conditionally controlled by the cascaded gating of the potential function; the model adopts an asynchronous parallel pipeline multi-channel structure, and each sub-layer independently performs belief propagation iteration; local asymmetric gating loop units are used within each channel to ensure the real-time performance of forward computation; through deep cascaded correction of multi-level potential functions, the model outputs the optimal spatiotemporal roaming path after the global energy function converges.

6. The method for dynamically generating roaming paths according to claim 5, characterized in that, The state entropy increase dynamic constraint function is used to quantify the uncertainty of obstacle trajectory distribution. The higher the entropy value, the stronger the constraint, forcing path points to converge to low-entropy regions. The group cooperative collision avoidance behavior function models the repulsive force between multiple agents as a potential field function. The intensity of the repulsive force is jointly determined by relative distance and relative velocity. The relative distance threshold is determined by repeatedly conducting obstacle avoidance success rate experiments in obstacle scenarios with different densities and performing piecewise regression analysis on the experimental data. The number of belief propagation iterations is determined by multiple rounds of comparative experiments on convergence speed and computation delay.

7. The method for dynamically generating roaming paths according to claim 6, characterized in that, The steps for establishing the training dataset for the Conditional Random Field Deep Extended Spatiotemporal Graph Evolution Model include: collecting 3D scene simulation data containing randomly moving obstacles of different densities, recording the rover pose sequence, the spatiotemporal trajectory of the obstacles, and the corresponding collision-free reference path, constructing a triplet dataset, and augmenting the data by applying random perturbations to the existing trajectories for any deficiencies; performing graph-structured preprocessing on the original trajectory data under the topological graph sparse adjacency propagation mechanism to generate node feature matrices and adjacency matrix pairs; the training steps include: using the composite energy function as the basic term of the loss function, combined with the path smoothness regularization term, and using an asynchronous parallel gradient accumulation strategy to update the network parameters of each subgraph channel, triggering early stopping when the energy function on the validation set does not decrease for multiple consecutive rounds.

8. The method for dynamically generating roaming paths according to claim 7, characterized in that, The specific structure of the linearized axial separation attention model for group cooperative collision avoidance behavior is as follows: The input layer projects multi-source heterogeneous inputs to a unified low-dimensional latent space through a shared latent space low-dimensional alignment operator; the bottom core adopts a linear attention approximation mechanism and an axial separation attention mechanism to decouple the traditional fully connected attention matrix into three one-dimensional attention calculation operators along the X-axis, Y-axis, and Z-axis, with each operator introducing a local sliding window; the behavioral repulsion force function of group cooperative collision avoidance behavior is embedded in the linear approximation operator as a weight scaling factor. Information flow is controlled in real time by cascaded gating attention; long-span residual dense jump matrix spans multiple axial attention layers; The output layer generates a smooth 3D roaming path point set tensor.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform a method for dynamically generating a roaming path in a three-dimensional scene as described in any one of claims 1-8.

10. A dynamic path generation system for a three-dimensional scene, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images