An artificial intelligence-based automatic driving path planning method for narrow space

Abstract

The application discloses a kind of narrow space automatic driving path planning method based on artificial intelligence, comprising the following steps: S1, laser radar point cloud and image data are collected, semantic segmentation and clustering are executed, and topological graph is established;S2, topological graph node is mapped as impulse coupling neural network neuron, and the double-domain coupling mechanism including structure coupling term and state coupling term is constructed;S3, path direction vector is generated, path intention gating factor is structured, and coupling input value is dynamically adjusted;S4, neuron membrane potential is periodically updated, and discharge threshold is adaptively adjusted;S5, the first discharge node number and time are recorded, and path activation sequence is generated according to discharge order;S6, connectivity check and curvature smoothing are performed on path activation sequence, and continuous trajectory is generated;S7, trajectory is converted into control command, and vehicle is driven to execute path tracking.The application realizes the high-precision, high-connectivity and high-control stability output of path planning in narrow space.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to an artificial intelligence-based method for autonomous driving path planning in narrow spaces. Background Technology

[0002] In autonomous driving technology, path planning systems for structured roads have been deployed in engineering. These systems mainly rely on rule-based maps, global search algorithms, and sensor feedback to generate paths and track control. Common algorithms such as Dijkstra's algorithm and improved heuristic search have high stability in rule-based environments. However, in areas with narrow spaces, limited access, and densely distributed obstacles, the path search process is prone to planning redundancy, path discontinuity, or abrupt curvature changes. This limits their effectiveness in application scenarios such as underground parking garages, tunnels, and warehouse passages. As the demand for path generation in complex geometric environments increases, grid maps built on the basis of single structural cost or distance cost are difficult to meet the accuracy requirements of dynamic adjustment and topological representation. This results in control commands not being executed smoothly, and vehicles are prone to swaying and yaw.

[0003] LiDAR and image fusion perception systems have been widely used in autonomous driving environment modeling. Passable areas can be extracted through semantic segmentation and point cloud clustering. However, most traditional mapping methods do not transform spatial connectivity into topological structures for path reasoning, but only generate surface occupancy representations or cost function graphs, lacking support for deep path propagation mechanisms. In recent years, deep learning models have gained attention in path reasoning methods, including path prediction and end-to-end trajectory generation based on neural networks. However, such methods rely on large-scale labeled samples and have limited ability to extrapolate structures. In highly constrained and narrow spaces, they are prone to problems such as path instability and lack of physical interpretability of reasoning results.

[0004] Some studies based on biological neural models have attempted to introduce spiking-coupled neural networks (SCLs) to perform path activation inference. These methods simulate neuronal firing and coupling propagation mechanisms to complete the path selection process, possessing a certain degree of temporal control and local response capability. However, existing SCLs do not incorporate the spatial state of nodes in their coupling structure design. When activation paths are set based on connection weights in a graph structure, the probability of neuronal activation is low and path activation capability is insufficient when bandwidth is narrow or there is high-density interference. The firing mechanism parameters are not dynamically adjusted according to actual spatial characteristics, leading to signal transmission obstruction or firing redundancy, affecting path connectivity and effectiveness. Existing firing processes do not distinguish between changes in the spatial environment and cannot achieve active control at channel contraction or locally complex nodes.

[0005] Most path activation results are generated by a fixed-time sorting method, without introducing structural connectivity verification and geometric curvature smoothing mechanisms. The path trajectory suffers from abrupt angle changes and control non-executability issues. In the path tracking control process, some methods directly fit the node sequence as the control output without considering the dynamic adjustment requirements of the path, which can easily lead to trajectory deviation and target error amplification. Existing technologies have failed to achieve coordinated optimization at the three levels of structural representation, state modeling and dynamic control, and lack highly reliable path planning methods suitable for autonomous driving environments in narrow spaces.

[0006] Therefore, how to provide an AI-based method for autonomous driving path planning in narrow spaces is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a path planning method for autonomous driving in narrow spaces based on pulse-coupled neural networks and a dual-domain coupling mechanism. This invention fully integrates LiDAR and image data, constructs an environmental topology map through semantic segmentation and spatial clustering, and uses structural coupling terms and state coupling terms to jointly drive the firing process of path neurons, thereby realizing the dynamic activation and optimized generation of paths. It has the advantages of strong path connectivity, high environmental adaptability, and smooth and controllable control trajectory.

[0008] An artificial intelligence-based path planning method for autonomous driving in narrow spaces according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect LiDAR point cloud and image data of autonomous vehicles in narrow spaces, perform semantic segmentation and clustering, extract passable areas and discretize them into spatial nodes, and establish a topology map;

[0010] S2. Using each node in the topology graph as a neuron of the pulse-coupled neural network, initialize the membrane potential, connection weights, coupling input values ​​and discharge thresholds, and construct a dual-domain coupling mechanism consisting of structural coupling terms and state coupling terms. The structural coupling terms are calculated based on the spatial connection relationship between nodes, and the state coupling terms are generated based on the passage width and obstacle risk level at the node.

[0011] S3. Calculate the path direction vector from the vehicle's current position to the target position, construct path intent gating factors for each node, and dynamically adjust the input value of the structural coupling term based on the projection result of the path direction vector.

[0012] S4. In each propagation cycle, update the neuron membrane potential. When the membrane potential exceeds the current threshold, trigger the discharge and transmit pulse signals to adjacent nodes. At the same time, adaptively reduce the discharge threshold according to the degree of node spatial compression.

[0013] S5. Record the number and time of the first discharge node, and connect adjacent nodes in ascending order of discharge time to generate a path activation sequence;

[0014] S6. Perform connectivity verification and curvature smoothing on the path activation sequence to generate a continuous and controllable path trajectory;

[0015] S7. Decompose the path trajectory into speed and steering commands at control time steps, and input them into the vehicle controller to execute path tracking.

[0016] Optionally, the feature is that S1 specifically includes: projecting the lidar point cloud onto a two-dimensional plane, generating a rasterized environment representation by combining the segmentation results of the image data, in the rasterized environment representation, selecting cells that meet the ground flatness threshold and obstacle confidence threshold as candidate passage areas, taking the center point of the candidate passage area as a spatial node, establishing node adjacency relationships based on the spatial distance between nodes, line-of-sight accessibility and surrounding obstacle density, organizing all spatial nodes and node adjacency relationships into a topology graph, wherein each node in the topology graph contains location coordinates and local environmental attributes, and each edge represents a feasible passage spatial connection relationship between nodes.

[0017] Optionally, the characteristic of step S2 is that it specifically includes:

[0018] S21. Set each spatial node in the topology graph as a neuron unit of a pulse-coupled neural network, and for each neuron... Initialize membrane potential Discharge threshold Discharge state Total Coupled Input Value and the matrix of connection weights ,in, Represents a node To the node The structural connection weights are set to zero initially.

[0019] S22, For each neuron The process of constructing a structurally coupled input includes:

[0020] S221, Acquiring Nodes The set of adjacent nodes in the topological connection S222, For each adjacent node Read the current discharge state Obtain spatial coordinates compute nodes With nodes Euclidean distance ,in, Indicate norm regularization; S223, calculate the distance perturbation factor based on the distance. ,in, It is an exponential function. For perturbation scale coefficients; S224, read node Obstacle risk level With passage width Calculate the risk modulation function:

[0021] ;

[0022] in, , To adjust the parameters; S225, based on the distance discharge state, distance perturbation factor, and risk modulation function information, the structural coupling input term is calculated as follows:

[0023] ;

[0024] in, This is the disturbance enhancement factor;

[0025] S23, For each neuron The execution of the state-coupled input construction process specifically includes:

[0026] S231, Statistical Nodes Historical number of passes Record the number of disturbance events per unit time. Let the disturbance suppression factor be... ,in, The disturbance offset constant; S232, obtain the node Direction vector pointing to adjacent nodes , and the target direction vector Compare and calculate the consistency of path direction. ;

[0027] S233, The calculation of the state coupling input term is:

[0028] ;

[0029] in, , These are the state coupling control coefficients;

[0030] S24. Add the structural coupling input term and the state coupling input term to obtain the node. At any moment Total coupling input value .

[0031] Optionally, the characteristic of step S3 is that it specifically includes:

[0032] S31. Obtain the current neuron Spatial position coordinates With target position coordinates ;

[0033] S32, For each neighboring neuron Obtain spatial coordinates Current cycle discharge state Obstacle risk modulation function value Path consistency metrics Calculate the path intent gating factor:

[0034] ;

[0035] in, The nonlinear adjustment coefficient representing the path gating factor;

[0036] S33, Set each adjacent node Structural coupling input terms Multiply by the corresponding path intent gating factor Complete the structural coupling terms The weighted correction yields the gated structure coupled input term. ;

[0037] S34. Couple the gating structure with the input terms. Input terms coupled with the calculated state Adding together generates neurons. The new total coupled input value for the current period .

[0038] Optionally, the characteristic of step S4 is that it specifically includes:

[0039] S41. In each propagation cycle, the receiving neuron Total coupled input value after gating correction ;

[0040] S42, Neuron membrane potential in the previous cycle The attenuation is performed according to the set attenuation ratio and coupled with the gating input of the current cycle. Adding them together yields the updated membrane potential. The attenuation ratio is used to control the decreasing trend of the membrane potential over time;

[0041] S43, when the updated membrane potential satisfy At that time, neurons Trigger discharge, set discharge signal Reset the membrane potential to If the conditions are not met, then set ;

[0042] S44. Based on the image data in S1, locate the neurons. The corresponding image region is used to count the number of pixels labeled as passable categories within the region, based on the topological graph structure, to count the nodes. The set of adjacent nodes is used to calculate the average connectivity. A spatial compression index is constructed based on the number of pixels in each passable category and the average connectivity. This index reflects the degree of passage restriction around a node. Then, the current discharge threshold is adjusted based on the spatial compression index. Perform adaptive updates to reduce the discharge threshold in narrow regions.

[0043] Optionally, the feature is that in S44, the neurons... The adaptive update process for the discharge threshold includes:

[0044] The number of pixels for each passable category is counted, and combined with the average connectivity of adjacent node sets in the topology graph, the spatial compression index for the current cycle is calculated, based on the discharge threshold and adjustment factor of the previous cycle. The product of the current cycle spatial compression index and the current cycle spatial compression index is used to generate the discharge threshold for the next cycle, wherein the adjustment factor A real number between 0 and 1, used to control the sensitivity of threshold changes;

[0045] The update strategy is divided into the following three categories based on the node space compression index: When the node is in a structurally stable and environmentally unobstructed area, an adjustment factor is set. The value is between 0.05 and 0.2, corresponding to a decrease in the discharge threshold controlled within the range of 5% to 15% compared to the previous cycle; when the node is in an area with local obstacles or obstructed path traffic capacity, the adjustment factor is set. The value is 0.3 to 0.6, corresponding to a decrease in the discharge threshold of 20% to 45%; when the node is in a dynamic obstacle environment, the adjustment factor is set. The value ranges from 0.7 to 0.9, corresponding to a decrease in the discharge threshold of 50% to 70%.

[0046] Optionally, the characteristic of step S5 is that it specifically includes:

[0047] S51. During each propagation cycle, monitor the discharge status of all nodes in the topology graph, record the moment when each node first triggers the discharge signal, and combine the node's corresponding number with the discharge time to form a discharge tag tuple.

[0048] S52. Sort all discharge marker tuples in ascending order of discharge time to construct a preliminary discharge node sequence;

[0049] S53. Based on the adjacency relationship in the topology graph, determine whether there is connectivity between adjacent nodes in the initial discharge node sequence. If there is a connecting edge, retain the connection order. If there is no connecting edge, backtrack to the previous node with the nearest connecting path and reconstruct the local connected subsequence until a complete path is formed.

[0050] S54. Concatenate all subsequences of nodes arranged in ascending order of time and possessing topological connectivity into a path activation sequence.

[0051] Optionally, the characteristic of step S6 is that it specifically includes:

[0052] S61. Input the generated path activation sequence into the topology graph verification module, and traverse the adjacent node pairs in the sequence. Determine whether there is an edge connecting adjacent node pairs in the topology graph. If there is an edge, retain the node pair. If there is no edge, find the shortest passable path from the original topology graph and insert it between the node pairs to complete the path segment repair.

[0053] S62. Denote the node sequence after connectivity verification as follows: ,based on continuous ternary node set Calculate the turning angle at each intermediate node. ,when Less than the set smoothing threshold At that time, a curvature smoothing operation is performed, the curvature smoothing operation including: using the current node Centered on a single point, a five-point weighted sliding window is used to adjust the centroid of the coordinates of the preceding and following path nodes, replacing the original node coordinates with the smoothed intermediate values ​​to generate a smooth path sequence. ;

[0054] S64. Output the final smoothed path sequence. , which serves as the input trajectory for continuous path tracking of vehicles.

[0055] The beneficial effects of this invention are:

[0056] Achieving high-precision modeling and topological representation of traversable regions in complex and narrow environments: This invention is based on the fusion of LiDAR point cloud and image data. Through semantic segmentation, traversable regions are extracted and discretized into spatial nodes, constructing a physically connected environmental topology map, providing a stable structural foundation for subsequent path reasoning. Compared to traditional grid or cost map methods, this invention has significant advantages in terms of spatial representation accuracy and connectivity integrity.

[0057] Enhance the environmental adaptability and dynamic adjustment capability of path activation inference: By introducing a pulse-coupled neural network model and constructing a dual-domain coupling mechanism that includes structural coupling terms and state coupling terms, it can simultaneously perceive the connection relationship between nodes and the state of the passage environment, realize the directional control of the coupled input value and the adaptive update of the discharge threshold, improve the effectiveness of path excitation in high compression regions and dense obstacle regions, and avoid path interruption and information propagation lag.

[0058] This invention generates a control path trajectory with strong connectivity and smooth curvature: By recording the time series of discharge nodes to construct a path activation sequence, performing topological connectivity verification and curvature smoothing processing, a continuous and controllable path trajectory is output. This trajectory is further decomposed into speed and steering commands, driving the vehicle to stably track in confined spaces. Compared to existing methods, this effectively reduces the risk of trajectory deviation and improves trajectory executability and control stability. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is an overall flowchart of an artificial intelligence-based autonomous driving path planning method in narrow spaces proposed in this invention.

[0061] Figure 2 This is a schematic diagram illustrating the traversable area extraction and topology map construction of an artificial intelligence-based autonomous driving path planning method in narrow spaces proposed in this invention.

[0062] Figure 3 This diagram illustrates the path activation sequence generation and trajectory smoothing process of an artificial intelligence-based autonomous driving path planning method for narrow spaces proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1-3 An artificial intelligence-based method for autonomous driving path planning in narrow spaces includes the following steps:

[0065] S1. Collect LiDAR point cloud and image data of autonomous vehicles in narrow spaces, perform semantic segmentation and clustering, extract passable areas and discretize them into spatial nodes, and establish a topology map;

[0066] S2. Using each node in the topology graph as a neuron of the pulse-coupled neural network, initialize the membrane potential, connection weights, coupling input values ​​and discharge thresholds, and construct a dual-domain coupling mechanism consisting of structural coupling terms and state coupling terms. The structural coupling terms are calculated based on the spatial connection relationship between nodes, and the state coupling terms are generated based on the passage width and obstacle risk level at the node.

[0067] S3. Calculate the path direction vector from the vehicle's current position to the target position, construct path intent gating factors for each node, and dynamically adjust the input value of the structural coupling term based on the projection result of the path direction vector.

[0068] S4. In each propagation cycle, update the neuron membrane potential. When the membrane potential exceeds the current threshold, trigger the discharge and transmit pulse signals to adjacent nodes. At the same time, adaptively reduce the discharge threshold according to the degree of node spatial compression.

[0069] S5. Record the number and time of the first discharge node, and connect adjacent nodes in ascending order of discharge time to generate a path activation sequence;

[0070] S6. Perform connectivity verification and curvature smoothing on the path activation sequence to generate a continuous and controllable path trajectory;

[0071] S7. Decompose the path trajectory into speed and steering commands at control time steps, and input them into the vehicle controller to execute path tracking.

[0072] This invention proposes a path planning method for autonomous driving in narrow spaces based on a pulse-coupled neural network and a dual-domain coupling mechanism. It integrates two types of spatial information: structural connectivity and state perception. Coupled driving firing is achieved at the node-level neurons. Compared with traditional A* algorithm and simple graph learning path planning methods, this invention constructs path activation sequences by simulating the firing process of biological neurons. It has stronger local environment adaptability and path connectivity guarantee capabilities, and can significantly improve the robustness and real-time performance of path generation in complex narrow channels.

[0073] In this embodiment, the characteristic is that S1 specifically includes: projecting the lidar point cloud onto a two-dimensional plane, generating a rasterized environment representation by combining the segmentation results of the image data, selecting cells that meet the ground flatness threshold and obstacle confidence threshold as candidate passage areas in the rasterized environment representation, using the center point of the candidate passage area as a spatial node, establishing node adjacency relationships based on the spatial distance between nodes, line-of-sight accessibility and surrounding obstacle density, and organizing all spatial nodes and node adjacency relationships into a topology graph, wherein each node in the topology graph contains location coordinates and local environmental attributes, and each edge represents a feasible passage spatial connection relationship between nodes.

[0074] In the process of extracting traversable regions and constructing topology graphs, this invention introduces a multimodal fusion method based on semantic segmentation and clustering. After discretizing the regions into spatial nodes, it constructs bidirectional connected edge relationships to generate a structural topology graph that can be used for neural model processing. Compared with existing path modeling methods based on grid graphs or regular maps, the topology graph constructed by this invention is more concise and expressive, and can provide a clearer graph structure foundation for subsequent neuronal structure modeling and coupling mechanism construction.

[0075] In this embodiment, the characteristic is that S2 specifically includes:

[0076] S21. Set each spatial node in the topology graph as a neuron unit of a pulse-coupled neural network, and for each neuron... Initialize membrane potential Discharge threshold Discharge state Total Coupled Input Value and the matrix of connection weights ,in, Represents a node To the node The structural connection weights are set to zero initially.

[0077] S22, For each neuron The process of constructing a structurally coupled input includes:

[0078] S221, Acquiring Nodes The set of adjacent nodes in the topological connection S222, For each adjacent node Read the current discharge state Obtain spatial coordinates compute nodes With nodes Euclidean distance ,in, Indicate norm regularization; S223, calculate the distance perturbation factor based on the distance. ,in, It is an exponential function. For perturbation scale coefficients; S224, read node Obstacle risk level With passage width Calculate the risk modulation function:

[0079] ;

[0080] in, , To adjust the parameters; S225, based on the distance discharge state, distance perturbation factor, and risk modulation function information, the structural coupling input term is calculated as follows:

[0081] ;

[0082] in, This is the disturbance enhancement factor;

[0083] S23, For each neuron The execution of the state-coupled input construction process specifically includes:

[0084] S231, Statistical Nodes Historical number of passes Record the number of disturbance events per unit time. Let the disturbance suppression factor be... ,in, The disturbance offset constant; S232, obtain the node Direction vector pointing to adjacent nodes , and the target direction vector Compare and calculate the consistency of path direction. ;

[0085] S233, The calculation of the state coupling input term is:

[0086] ;

[0087] in, , These are the state coupling control coefficients;

[0088] S24. Add the structural coupling input term and the state coupling input term to obtain the node. At any moment Total coupling input value .

[0089] This invention creatively introduces a dual-domain coupling mechanism into the pulse-coupled neural network model, fusing spatial connectivity structures with node traffic status indicators. By constructing structural coupling terms and state coupling terms to reflect spatial connectivity and local traffic risk respectively, coupled input values ​​are formed at the network input side. Compared to traditional fixed coupling strategies, this invention can dynamically regulate neuron state updates during path propagation, effectively improving the environmental sensitivity and dynamic adjustment capability of path propagation behavior.

[0090] In this embodiment, the characteristic is that S3 specifically includes:

[0091] S31. Obtain the current neuron Spatial position coordinates With target position coordinates ;

[0092] S32, For each neighboring neuron Obtain spatial coordinates Current cycle discharge state Obstacle risk modulation function value Path consistency metrics Calculate the path intent gating factor:

[0093] ;

[0094] in, The nonlinear adjustment coefficient representing the path gating factor;

[0095] S33, Set each adjacent node Structural coupling input terms Multiply by the corresponding path intent gating factor Complete the structural coupling terms The weighted correction yields the gated structure coupled input term. ;

[0096] S34. Couple the gating structure with the input terms. Input terms coupled with the calculated state Adding together generates neurons. The new total coupled input value for the current period .

[0097] This invention designs a path intent gating factor mechanism for the path direction guidance problem. It maps the direction vector from the vehicle's current position to the target position to the spatial position direction of neuron nodes, and adjusts the coupling input strength in the structural coupling term accordingly, ensuring that neuron firing propagation more closely aligns with the global target path direction. Compared to path propagation methods that rely solely on local node states, this invention improves the goal consistency of path planning within the guidance mechanism, thereby enhancing the global optimality and practical feasibility of the planned path.

[0098] In this embodiment, the characteristic is that S4 specifically includes:

[0099] S41. In each propagation cycle, the receiving neuron Total coupled input value after gating correction ;

[0100] S42, Neuron membrane potential in the previous cycle The attenuation is performed according to the set attenuation ratio and coupled with the gating input of the current cycle. Adding them together yields the updated membrane potential. The attenuation ratio is used to control the decreasing trend of the membrane potential over time;

[0101] S43, when the updated membrane potential satisfy At that time, neurons Trigger discharge, set discharge signal Reset the membrane potential to If the conditions are not met, then set ;

[0102] S44. Based on the image data in S1, locate the neurons. The corresponding image region is used to count the number of pixels labeled as passable categories within the region, based on the topological graph structure, to count the nodes. The set of adjacent nodes is used to calculate the average connectivity. A spatial compression index is constructed based on the number of pixels in each passable category and the average connectivity. This index reflects the degree of passage restriction around a node. Then, the current discharge threshold is adjusted based on the spatial compression index. Perform adaptive updates to reduce the discharge threshold in narrow regions.

[0103] This invention introduces a node spatial compression index to adaptively adjust the firing threshold in the pulse-coupled neuron membrane potential update mechanism. The threshold is dynamically reduced based on environmental compression during firing, thereby increasing the sensitivity of path activation in narrow regions. This mechanism differs from fixed-threshold models, automatically adjusting path propagation speed and node activity according to path complexity and spatial conditions, achieving efficient path advancement and stable firing control in complex channels.

[0104] In this embodiment, the characteristic is that in S44, the neurons... The adaptive update process for the discharge threshold includes:

[0105] The number of pixels for each passable category is counted, and combined with the average connectivity of adjacent node sets in the topology graph, the spatial compression index for the current cycle is calculated, based on the discharge threshold and adjustment factor of the previous cycle. The product of the current cycle spatial compression index and the current cycle spatial compression index is used to generate the discharge threshold for the next cycle, wherein the adjustment factor A real number between 0 and 1, used to control the sensitivity of threshold changes;

[0106] The update strategy is divided into the following three categories based on the node space compression index: When the node is in a structurally stable and environmentally unobstructed area, an adjustment factor is set. The value is between 0.05 and 0.2, corresponding to a decrease in the discharge threshold controlled within the range of 5% to 15% compared to the previous cycle; when the node is in an area with local obstacles or obstructed path traffic capacity, the adjustment factor is set. The value is 0.3 to 0.6, corresponding to a decrease in the discharge threshold of 20% to 45%; when the node is in a dynamic obstacle environment, the adjustment factor is set. The value ranges from 0.7 to 0.9, corresponding to a decrease in the discharge threshold of 50% to 70%.

[0107] This invention proposes a classification-based adjustment strategy for the adaptive discharge threshold update mechanism, which adapts to changes in discharge rhythm. Depending on the threshold decrease rate and fluctuation amplitude within a cycle, it employs various adjustment methods such as gradual descent, enhanced buffering, and accelerated descent to achieve differentiated responses to varying complexity in different passage regions. Unlike existing single-descent or fixed-threshold mechanisms, this invention enhances the flexibility and intelligence of threshold control through a classification-based response mechanism, thereby improving the stability and efficiency of the path activation process.

[0108] In this embodiment, the characteristic is that S5 specifically includes:

[0109] S51. During each propagation cycle, monitor the discharge status of all nodes in the topology graph, record the moment when each node first triggers the discharge signal, and combine the node's corresponding number with the discharge time to form a discharge tag tuple.

[0110] S52. Sort all discharge marker tuples in ascending order of discharge time to construct a preliminary discharge node sequence;

[0111] S53. Based on the adjacency relationship in the topology graph, determine whether there is connectivity between adjacent nodes in the initial discharge node sequence. If there is a connecting edge, retain the connection order. If there is no connecting edge, backtrack to the previous node with the nearest connecting path and reconstruct the local connected subsequence until a complete path is formed.

[0112] S54. Concatenate all subsequences of nodes arranged in ascending order of time and possessing topological connectivity into a path activation sequence.

[0113] This invention proposes a path activation sequence generation mechanism based on discharge time recording, and integrates topology structure checking and chain break repair strategies to ensure the complete connectivity of discharge paths within the spatial graph structure. Furthermore, a curvature smoothing operation is introduced in the sequence post-processing stage to further improve the geometric continuity of the paths. Compared to traditional path extraction methods based on cost accumulation or gradient diffusion, the method of this invention outperforms traditional methods in terms of structural reachability and geometric rationality, improving the realistic controllability of paths in physical scenarios.

[0114] In this embodiment, the characteristic is that S6 specifically includes:

[0115] S61. Input the generated path activation sequence into the topology graph verification module, and traverse the adjacent node pairs in the sequence. Determine whether there is an edge connecting adjacent node pairs in the topology graph. If there is an edge, retain the node pair. If there is no edge, find the shortest passable path from the original topology graph and insert it between the node pairs to complete the path segment repair.

[0116] S62. Denote the node sequence after connectivity verification as follows: ,based on continuous ternary node set Calculate the turning angle at each intermediate node. ,when Less than the set smoothing threshold At that time, a curvature smoothing operation is performed, the curvature smoothing operation including: using the current node Centered on a single point, a five-point weighted sliding window is used to adjust the centroid of the coordinates of the preceding and following path nodes, replacing the original node coordinates with the smoothed intermediate values ​​to generate a smooth path sequence. ;

[0117] S64. Output the final smoothed path sequence. , which serves as the input trajectory for continuous path tracking of vehicles.

[0118] This invention introduces a turning angle detection method based on a three-node sliding window during the path smoothing stage, and performs dynamic compensation and smoothing interpolation operations at locations of abrupt curvature changes, ultimately outputting a controllable trajectory path suitable for tracking. Compared to global smoothing algorithms based on spline interpolation or Bezier fitting, this method has higher local sensitivity and boundary preservation capabilities, making it particularly suitable for path planning tasks involving frequent turns in narrow spaces, thus improving path executability and control response stability.

[0119] Example 1:

[0120] To verify the feasibility of this invention in practice, it was applied to an autonomous driving task of a low-speed unmanned transport vehicle in a narrow underground passage. This scenario involved a long, enclosed environment with multiple curves, numerous obstacles, uneven structural compression, and a minimum passage width of less than 1.2 meters. The passage contained various obstacles, including stacked debris, low beams, and structural interference from wall corners. Furthermore, the different ground materials and reflective properties caused perceptual interference to the lidar and image recognition systems, making it a typical high-risk and challenging environment for path planning.

[0121] In this scenario, the vehicle is equipped with a 16-line LiDAR and a binocular camera. The front-end perception module collects LiDAR point clouds and image information from the passageway, performs joint semantic segmentation, extracts the passage area, and performs cluster analysis, discretizing the connected region into several spatial nodes, each node representing a local passageway. A topology graph model is constructed based on nodes, and spatial connections are established between nodes to form connected edges. Then, the topology graph is mapped to a pulse-coupled neuron structure in a neural network model. Initial membrane potential, connection weights, discharge thresholds, and coupling inputs are set for each node. The structural coupling term is calculated based on the connection weights between nodes, while the state coupling term is constructed based on the passage width at the node (measured value range 1.1m~2.6m) and the obstacle risk level (from level 0 to 3), forming a dual-domain coupled input structure.

[0122] During navigation, a path direction vector is generated between the vehicle's starting point and the target point. The system calculates the projection direction for all neuron nodes, constructs a path intent gating factor, and dynamically adjusts the coupling input strength, enhancing the coupling input of nodes aligned with the target path and weakening the input of nodes deviating from the path direction. Membrane potentials are updated each cycle. If a node's membrane potential exceeds the current threshold, it immediately discharges, and the system automatically lowers the subsequent discharge thresholds based on the compression ratio of the node's spatial location. The system records the time and sequence of all first discharges, generating a path activation sequence. This sequence then undergoes a topological connectivity check. If no valid edge connects any two nodes, the system automatically inserts an alternative path from the topology graph to ensure path integrity. Finally, a three-node curvature smoothing process is performed on the path, removing nodes with turning angles greater than 70 degrees and frequent continuous turns, optimizing it into a continuous and controllable trajectory, and generating speed and steering control commands.

[0123] In this scenario, we conducted experiments comparing the traditional A* algorithm, end-to-end path prediction methods based on graph convolutional networks, and the path activation mechanism proposed in this invention. The experiment involved 50 complete passage tasks, each with different start and end point combinations. Measurements included path connectivity, average trajectory curvature, obstacle avoidance success rate, control command smoothness, and final trajectory deviation error. Our results show that in complex and narrow areas, the proposed method exhibits stronger robustness in terms of path connectivity, control stability, and local adaptability. The experimental results are summarized in the table below.

[0124] Table 1: Performance Comparison Data of Different Methods in Narrow Space Path Planning Tasks

[0125]

[0126] This experiment compared three typical path planning methods: the traditional A algorithm, the graph convolutional neural network (GCN) path prediction method, and the method proposed in this invention based on a pulse-coupled neural network and a dual-domain coupling mechanism. In terms of path connectivity, the proposed method performed best, consistently maintaining above 96% in multiple rounds of experiments, reaching a maximum of 97.5%, significantly outperforming the A algorithm (average approximately 86.7%) and the graph convolutional method (average approximately 90.5%). This result demonstrates that the proposed method possesses stronger path feasibility assurance capabilities in complex structure compression environments.

[0127] From the perspective of maximum trajectory curvature, the traditional A* method produces a relatively large trajectory curvature, with an average maximum value of about 78 degrees. Graph convolution methods smooth the trajectory through end-to-end learning, reducing the maximum curvature to about 60 degrees. However, the method of this invention, due to its discharge path sorting and three-node sliding window smoothing strategy, effectively reduces the steepness of path turns, achieving an average maximum curvature of 46 degrees. This results in better trajectory continuity and controllability in narrow environments, which is beneficial for path tracking by low-speed autonomous vehicles.

[0128] In terms of obstacle avoidance success rate, the average obstacle avoidance rate of this invention is 94.8%, which is much higher than the 82.3% of the A* method and the 87.8% of the graph convolution method. This is due to the fact that this invention introduces joint modeling of passage width and obstacle level in the state coupling term, which makes path activation more selective and anti-interference, thereby effectively avoiding obstacle areas and improving obstacle avoidance performance.

[0129] Control command volatility reflects the vehicle's stability during path execution. Data shows that the A* algorithm's control signal volatility is relatively large, averaging 0.43 rad / s, while the graph convolution method reduces it to 0.36 rad / s, and the method of this invention is only around 0.20 rad / s. A smoother control signal means greater vehicle stability during passage, smoother action adjustments, and significantly reduced jitter and errors caused by sudden attitude changes during path execution.

[0130] Regarding trajectory accuracy, the method of this invention has an average trajectory deviation of 0.17m, which is significantly better than the 0.33m of the graph convolution method and the 0.40m of the A* method. The lower deviation value verifies that the path activation sequence constructed by this method has higher physical reachability and geometric consistency, indicating that the discharge mechanism can improve execution accuracy while preserving path accessibility.

[0131] The above analysis results demonstrate that this invention outperforms existing technologies in several key performance dimensions, exhibiting stronger engineering practicality and environmental adaptability, and is particularly suitable for intelligent path planning scenarios in narrow and complex spaces. Through collaborative modeling of structural coupling and state coupling, as well as an adaptive adjustment mechanism for the discharge threshold, it can efficiently achieve an optimized closed loop for the entire process of path activation, planning, and control.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for autonomous driving path planning in narrow spaces based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect LiDAR point cloud and image data of autonomous vehicles in narrow spaces, perform semantic segmentation and clustering, extract passable areas and discretize them into spatial nodes, and establish a topology map; S2. Using each node in the topology graph as a neuron of the pulse-coupled neural network, initialize the membrane potential, connection weights, coupling input values ​​and discharge thresholds, and construct a dual-domain coupling mechanism consisting of structural coupling terms and state coupling terms. The structural coupling terms are calculated based on the spatial connection relationship between nodes, and the state coupling terms are generated based on the passage width and obstacle risk level at the node. S3. Calculate the path direction vector from the vehicle's current position to the target position, construct path intent gating factors for each node, and dynamically adjust the input value of the structural coupling term based on the projection result of the path direction vector. S4. In each propagation cycle, update the neuron membrane potential. When the membrane potential exceeds the current threshold, trigger the discharge and transmit pulse signals to adjacent nodes. At the same time, adaptively reduce the discharge threshold according to the degree of node spatial compression. S5. Record the number and time of the first discharge node, and connect adjacent nodes in ascending order of discharge time to generate a path activation sequence; S6. Perform connectivity verification and curvature smoothing on the path activation sequence to generate a continuous and controllable path trajectory; S7. Decompose the path trajectory into speed and steering commands at control time steps, and input them into the vehicle controller to execute path tracking.

2. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 1, characterized in that, S1 specifically includes: projecting the LiDAR point cloud onto a two-dimensional plane, generating a rasterized environment representation based on the segmentation results of the image data, selecting cells that meet the ground flatness threshold and obstacle confidence threshold as candidate passage areas in the rasterized environment representation, using the center point of the candidate passage area as a spatial node, establishing node adjacency relationships based on the spatial distance between nodes, line-of-sight accessibility, and surrounding obstacle density, and organizing all spatial nodes and node adjacency relationships into a topology graph, wherein each node in the topology graph contains location coordinates and local environmental attributes, and each edge represents a feasible spatial connection relationship between nodes.

3. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 2, characterized in that, S2 specifically includes: S21. Set each spatial node in the topology graph as a neuron unit of a pulse-coupled neural network, and for each neuron... Initialize membrane potential Discharge threshold Discharge state Total Coupled Input Value and the matrix of connection weights ,in, Represents a node To the node The structural connection weights are set to zero initially. S22, For each neuron The process of constructing a structurally coupled input includes: S221, Acquiring Nodes The set of adjacent nodes in the topological connection S222, For each adjacent node Read the current discharge state Obtain spatial coordinates compute nodes With nodes Euclidean distance ,in, Indicate norm regularization; S223, calculate the distance perturbation factor based on the distance. ,in, It is an exponential function. For perturbation scale coefficients; S224, read node Obstacle risk level With passage width Calculate the risk modulation function: ； in, , To adjust the parameters; S225, based on the distance discharge state, distance perturbation factor, and risk modulation function information, the structural coupling input term is calculated as follows: ； in, This is the disturbance enhancement factor; S23, For each neuron The execution of the state-coupled input construction process specifically includes: S231, Statistical Nodes Historical number of passes Record the number of disturbance events per unit time. Let the disturbance suppression factor be... ,in, The disturbance offset constant; S232, obtain the node Direction vector pointing to adjacent nodes , and the target direction vector Compare and calculate the consistency of path direction. ; S233, The calculation of the state coupling input term is: ； in, , These are the state coupling control coefficients; S24. Add the structural coupling input term and the state coupling input term to obtain the node. At any moment Total coupling input value .

4. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 3, characterized in that, S3 specifically includes: S31. Obtain the current neuron Spatial position coordinates With target position coordinates ; S32, For each neighboring neuron Obtain spatial coordinates Current cycle discharge state Obstacle risk modulation function value Path consistency metrics Calculate the path intent gating factor: ； in, The nonlinear adjustment coefficient representing the path gating factor; S33, Set each adjacent node Structural coupling input terms Multiply by the corresponding path intent gating factor Complete the structural coupling terms The weighted correction yields the gated structure coupled input term. ; S34. Couple the gating structure with the input terms. Input terms coupled with the calculated state Adding together generates neurons. The new total coupled input value for the current period .

5. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 4, characterized in that, S4 specifically includes: S41. In each propagation cycle, the receiving neuron Total coupled input value after gating correction ; S42, Neuron membrane potential in the previous cycle The attenuation is performed according to the set attenuation ratio and coupled with the gating input of the current cycle. Adding them together yields the updated membrane potential. The attenuation ratio is used to control the decreasing trend of the membrane potential over time; S43, when the updated membrane potential satisfy At that time, neurons Trigger discharge, set discharge signal Reset the membrane potential to If the conditions are not met, then set ; S44. Based on the image data in S1, locate the neurons. The corresponding image region is used to count the number of pixels labeled as passable categories within the region, based on the topological graph structure, to count the nodes. The set of adjacent nodes is used to calculate the average connectivity. A spatial compression index is constructed based on the number of pixels in each passable category and the average connectivity. This index reflects the degree of passage restriction around a node. Then, the current discharge threshold is adjusted based on the spatial compression index. Perform adaptive updates to reduce the discharge threshold in narrow regions.

6. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 5, characterized in that, In S44, the neurons The adaptive update process for the discharge threshold includes: The number of pixels for each passable category is counted, and combined with the average connectivity of adjacent node sets in the topology graph, the spatial compression index for the current cycle is calculated, based on the discharge threshold and adjustment factor of the previous cycle. The product of the current cycle spatial compression index and the current cycle spatial compression index is used to generate the discharge threshold for the next cycle, wherein the adjustment factor A real number between 0 and 1, used to control the sensitivity of threshold changes; The update strategy is divided into the following three categories based on the node space compression index: When the node is in a structurally stable and environmentally unobstructed area, an adjustment factor is set. The value is between 0.05 and 0.2, corresponding to a decrease in the discharge threshold of 5% to 15% compared to the previous cycle; when the node is in an area with local obstacles or obstructed path traffic, the adjustment factor is set. The value is 0.3 to 0.6, corresponding to a decrease in the discharge threshold of 20% to 45%; when the node is in a dynamic obstacle environment, the adjustment factor is set. The value ranges from 0.7 to 0.9, corresponding to a decrease in the discharge threshold of 50% to 70%.

7. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 6, characterized in that, S5 specifically includes: S51. During each propagation cycle, monitor the discharge status of all nodes in the topology graph, record the moment when each node first triggers the discharge signal, and combine the node's corresponding number with the discharge time to form a discharge tag tuple. S52. Sort all discharge marker tuples in ascending order of discharge time to construct a preliminary discharge node sequence; S53. Based on the adjacency relationship in the topology graph, determine whether there is connectivity between adjacent nodes in the initial discharge node sequence. If there is a connecting edge, retain the connection order. If there is no connecting edge, backtrack to the previous node with the nearest connecting path and reconstruct the local connected subsequence until a complete path is formed. S54. Concatenate all subsequences of nodes arranged in ascending order of time and possessing topological connectivity into a path activation sequence.

8. The method for autonomous driving path planning in narrow spaces based on artificial intelligence according to claim 7, characterized in that, S6 specifically includes: S61. Input the generated path activation sequence into the topology graph verification module, and traverse the adjacent node pairs in the sequence. Determine whether there is an edge connecting adjacent node pairs in the topology graph. If there is an edge, retain the node pair. If there is no edge, find the shortest passable path from the original topology graph and insert it between the node pairs to complete the path segment repair. S62. Denote the node sequence after connectivity verification as follows: ,based on continuous ternary node set Calculate the turning angle at each intermediate node. ,when Less than the set smoothing threshold At that time, a curvature smoothing operation is performed, the curvature smoothing operation including: using the current node Centered on a single point, a five-point weighted sliding window is used to adjust the centroid of the coordinates of the preceding and following path nodes, replacing the original node coordinates with the smoothed intermediate values ​​to generate a smooth path sequence. ; S64. Output the final smoothed path sequence. , which serves as the input trajectory for continuous path tracking of vehicles.

Images