Deep Learning-Based High-Precision Positioning Cone Robot Path Planning System

Abstract

This invention discloses a high-precision localization path planning system for a road cone robot based on deep learning, comprising: a visual acquisition and preprocessing module for acquiring and preprocessing road scene data; a visual semantic recognition module for generating a set of scene semantic information; a multi-source localization fusion module for obtaining a sequence of robot localization results; a semantic map construction module for generating a road semantic map; a spatiotemporal energy constraint construction module for constructing a spatiotemporal energy constraint field based on the road semantic map; an energy-optimized path planning module for generating path planning results; a dynamic path correction module for correcting the path planning results; and a trajectory tracking control module for driving the road cone robot to run along a dynamic path. This invention employs deep learning and spatiotemporal energy optimization methods to achieve dynamic path planning for the road cone robot, possessing advantages such as high localization accuracy, intelligent obstacle avoidance, and strong adaptability.

Description

Technical Field

[0001] This invention relates to the field of path planning, and more particularly to a high-precision positioning path planning system for a road cone robot based on deep learning. Background Technology

[0002] Currently, semi-automatic or remote-controlled traffic cone deployment robots are commonly used in road construction and traffic management. Their path planning usually relies on fixed paths or preset trajectories based on static maps, failing to achieve real-time perception and adaptive adjustment of the dynamic environment. Existing systems are independent of path planning and environmental perception, often using traditional geometric mapping and rule-based obstacle avoidance algorithms. They lack unified modeling of lane structure, spatial relationships of obstacles, and semantics of deployment tasks, making path planning results prone to failure when the road environment changes.

[0003] Furthermore, although existing high-precision positioning systems can achieve basic position fusion using GNSS and IMU, their positioning results and visual semantic information lack dynamic collaborative constraints, making it impossible to reflect energy continuity and optimizability in path planning. Traditional planning algorithms are mostly based on shortest path search with fixed cost functions, lacking learnable global energy representation, making it difficult to achieve dynamic balance between road semantics, obstacle avoidance, and target deployment points, resulting in reduced deployment accuracy, insufficient path stability, and lagging obstacle avoidance strategies. Summary of the Invention

[0004] One objective of this invention is to propose a high-precision positioning path planning system for a road cone robot based on deep learning. This invention uses deep learning and spatiotemporal energy optimization methods to realize dynamic path planning for the road cone robot, which has the advantages of high positioning accuracy, intelligent obstacle avoidance, and strong adaptability.

[0005] A high-precision localization path planning system for a traffic cone robot based on deep learning, according to an embodiment of the present invention, includes:

[0006] The visual acquisition and preprocessing module is used to acquire visual image data, GNSS positioning data, RTK differential data and IMU inertial measurement data of road scenes, and perform preprocessing.

[0007] The visual semantic recognition module is used to input preprocessed visual image data into the visual recognition network, perform lane area segmentation, obstacle detection, road edge extraction and semantic element recognition on the road scene, and generate a set of scene semantic information.

[0008] The multi-source positioning fusion module is used to perform multi-source positioning fusion based on preprocessed GNSS positioning data, RTK differential data and IMU inertial measurement data to obtain a sequence of robot positioning results.

[0009] The semantic map construction module is used to perform spatiotemporal matching of scene semantic information sets with localization result sequences to generate road semantic maps;

[0010] The spatiotemporal energy constraint construction module is used to construct a spatiotemporal energy constraint field based on the road semantic map, mapping the lane centerline, obstacle boundary and road cone placement points as energy attraction source, repulsion source and target energy attraction source respectively, forming a spatiotemporal energy distribution structure that can be continuously optimized;

[0011] The energy optimization path planning module is used to input the spatiotemporal energy constraint field into the energy optimization network and generate path planning results based on the principle of energy minimization.

[0012] The dynamic path correction module is used to correct the path planning results based on the real-time updated set of scene semantic information and the sequence of positioning results, and generate dynamic path planning results.

[0013] The trajectory tracking control module is used to input the dynamic path planning results into the trajectory tracking controller, generate a continuous sequence of control commands, and drive the road cone robot to run along the dynamic path.

[0014] Optionally, the visual acquisition and preprocessing module includes:

[0015] Collect visual image data, GNSS positioning data, RTK differential data, and IMU inertial measurement data of road scenes;

[0016] Perform standardization operations on visual image data;

[0017] Perform formatting operations on GNSS positioning data, RTK differential data, and IMU inertial measurement data;

[0018] Time synchronization is performed on GNSS positioning data, RTK differential data and IMU inertial measurement data, and the timestamps of each data source are unified to a global time reference to form a time-synchronized multi-source positioning data sequence;

[0019] Spatial coordinate registration is performed on the time-synchronized multi-source positioning data sequence.

[0020] Optionally, the standardization operations include color equalization, brightness normalization, noise suppression, and lens distortion correction.

[0021] Optionally, the visual semantic recognition module includes:

[0022] The preprocessed visual image data is input into the visual recognition network to perform feature extraction and semantic encoding on the road scene, generating a visual feature map that includes spatial and semantic features;

[0023] In the lane region segmentation stage, pixel-by-pixel semantic classification is performed based on visual feature mapping to identify and divide the lane center region, lane boundary region and non-lane region in the road scene, generating a lane region segmentation result that includes lane region labels and pixel-level classification results.

[0024] In the obstacle detection stage, visual feature mapping is used to generate candidate regions and identify targets, and vehicles, pedestrians, traffic cones and fixed obstacles in the road scene are identified and located, and the obstacle detection results containing target category labels and boundary positions are output.

[0025] In the road edge extraction stage, the road edge pixel sequence is extracted based on the gradient change and geometric continuity information in the visual feature mapping, and the road edge feature set is generated by the curve fitting method.

[0026] In the semantic element recognition stage, the lane area segmentation results, obstacle detection results, and road edge feature sets are integrated and fused according to spatial correspondence to identify semantic elements of lane areas, road edges, and obstacles, generating a set of scene semantic information containing semantic labels, pixel-level segmentation results, and spatial coordinates.

[0027] Optionally, the multi-source positioning fusion module includes:

[0028] Calculate the robot’s initial position in the geographic coordinate system based on the preprocessed GNSS positioning data, and generate the GNSS initial positioning result.

[0029] The preprocessed RTK differential data is used to perform differential correction on the initial GNSS positioning results to obtain high-precision position coordinates;

[0030] The robot's attitude angles are calculated based on the preprocessed IMU inertial measurement data, resulting in attitude angles including roll angle, pitch angle and yaw angle;

[0031] Multi-source positioning fusion is performed on high-precision position coordinates and attitude angles. GNSS positioning data, RTK differential data and IMU inertial measurement data are fused and solved to generate positioning results that include robot three-dimensional coordinates and attitude angles.

[0032] The location results are organized in chronological order to form a location result sequence.

[0033] Optionally, the semantic map construction module includes:

[0034] The scene semantic information set and the localization result sequence are matched according to timestamps. The scene semantic information set is projected onto the global coordinate system to establish the spatiotemporal correspondence between the scene semantic information and the robot pose, and a global semantic mapping data set is generated.

[0035] In the global semantic mapping dataset, the spatial point set corresponding to the semantic labels of the lane area is structurally extracted to form lane structure information containing the lane centerline and lane boundary.

[0036] Clustering and spatial localization are performed on the spatial point set corresponding to the semantic tags of obstacles to determine the spatial position and boundary shape of the obstacles in the global coordinate system and generate obstacle position information;

[0037] Based on lane structure information and obstacle location information, drivable area extraction is performed. Spatial difference and boundary calculation are performed on the lane centerline point set, boundary point set and obstacle point set to determine the drivable area boundary and generate drivable area information.

[0038] The location information of road cones is determined from the information of drivable areas. The coordinate set of road cone locations is generated according to the preset spacing, orientation and safety distance conditions, with the center line of the lane as the reference.

[0039] By spatiotemporally fusing and structurally combining lane structure information, obstacle location information, drivable area information, and traffic cone placement information, a road semantic map containing lane structure, obstacle location, drivable area, and traffic cone placement information is generated.

[0040] Optionally, the spatiotemporal energy constraint construction module includes:

[0041] Based on the lane structure, obstacle location, drivable area and road cone layout information in the road semantic map, a spatiotemporal energy constraint field coordinate framework is established in the global coordinate system.

[0042] Weight parameters are defined according to the semantic element categories. The semantics of the lane area are mapped to the attraction weight parameters, the semantics of the obstacle are mapped to the repulsion weight parameters, and the semantics of the traffic cone placement points are mapped to the target attraction weight parameters, forming a set of semantic energy mapping parameters.

[0043] Based on the lane structure information, the geometric curvature and directional continuity of the lane centerline direction are calculated, and the attraction energy intensity is allocated to the lane center region to form a continuous energy attraction distribution band in the lane center direction, so that the energy maintains a smooth gradient change in the center region, which constitutes the main attraction path of the robot in the spatiotemporal energy constraint field.

[0044] Based on the obstacle's location information and geometric parameters, a local repulsion energy region is generated, and a repulsion energy rising zone is constructed around the obstacle, so that the energy gradually increases in the region close to the obstacle, forming a spatial repulsion zone;

[0045] Based on the location information of the road cones, a local energy attraction center is constructed in the global coordinate system, and a local energy minima are established at each road cone location, so that the energy forms a smoothly converging attraction potential trap at that location.

[0046] By integrating the energy distribution of lane attraction, the energy distribution of obstacle repulsion, and the energy distribution of traffic cone placement points, a continuous and differentiable spatiotemporal energy constraint field is established.

[0047] Optionally, the energy optimization path planning module includes:

[0048] The spatiotemporal energy constraint field is input into the energy optimization network, which performs path optimization calculations based on the energy distribution characteristic values, spatial location coordinates, and time index in the spatiotemporal energy constraint field.

[0049] Based on the energy distribution characteristics in the spatiotemporal energy constraint field, the energy difference between adjacent spatial coordinates is weighted and normalized to calculate the path energy cost.

[0050] The path node sequence is iteratively updated based on the principle of minimizing energy until the absolute value of the difference between the path energy cost of two consecutive iterations is less than the preset convergence threshold, thus generating the optimal path node set.

[0051] Based on the spatial coordinates of the energy-optimal path node set, curve fitting is performed to calculate the trajectory curvature and orientation angle of each path node, and a path node velocity distribution sequence is generated based on the time index.

[0052] The path node sequence and velocity distribution sequence are combined to output the path planning result, which includes the path node sequence, trajectory curvature, and velocity distribution sequence.

[0053] Optionally, the dynamic path correction module includes:

[0054] Receives a set of scene semantic information and a sequence of positioning results that are updated in real time;

[0055] The real-time scene semantic information set and the localization result sequence are compared with the current path planning result to identify the affected areas in the path;

[0056] The lane structure, obstacle location, and traffic cone placement information are updated based on the latest scene semantic information set, and the positions of affected path nodes are adjusted.

[0057] Based on the adjusted path node sequence, the path direction and node connection order are recalculated to generate the corrected path planning result;

[0058] The revised path planning results are matched and verified with the real-time positioning results to ensure path continuity and feasibility, and dynamic path planning results are generated.

[0059] The beneficial effects of this invention are:

[0060] The proposed high-precision positioning road cone robot path planning system based on deep learning achieves integrated collaboration of path planning, environmental perception, and positioning control through multi-source data fusion, semantic recognition, and energy optimization modeling. The system comprehensively utilizes visual image data, GNSS positioning data, RTK differential data, and IMU inertial measurement data. By generating multi-source fused data sequences through time synchronization and spatial coordinate registration, it ensures consistency and accuracy matching of different data sources in the spatiotemporal dimensions, providing high-quality input for subsequent semantic recognition and positioning fusion. Through a visual recognition network, it achieves lane area segmentation, obstacle detection, road edge extraction, and semantic element recognition. The system can generate a set of scene semantic information containing semantic labels, spatial coordinates, and pixel-level segmentation results, providing accurate environmental semantic descriptions for path construction.

[0061] This invention uses a multi-source positioning fusion module to jointly solve GNSS, RTK, and IMU data, generating a sequence of positioning results containing the robot's three-dimensional coordinates and attitude angles, achieving centimeter-level dynamic positioning accuracy. It then performs spatiotemporal matching of the scene semantic information set with the positioning result sequence to generate a road semantic map containing information on lane structure, obstacle locations, drivable areas, and road cone placement points. This provides a unified spatial representation for energy constraint modeling. Based on the road semantic map, a spatiotemporal energy constraint field is established, and the lane centerline, obstacle boundaries, and road cone placement points are mapped as attraction sources, repulsion sources, and target attraction sources, respectively, forming a continuously optimizable energy distribution structure that enables path planning to possess continuity and differentiability.

[0062] This invention further inputs a spatiotemporal energy constraint field into an energy optimization network, generates path planning results based on the principle of energy minimization, and ensures the smoothness and energy optimization of the path in the global scope. When road environment information or positioning data is updated in real time, the system can correct the path planning results in real time according to the new scene semantic information set and positioning result sequence, and output dynamic path planning results to ensure that the path is continuous and feasible. The trajectory tracking control module converts the dynamic path planning results into continuous control commands to drive the road cone robot to run smoothly along the planned path. Attached Figure Description

[0063] 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:

[0064] Figure 1 This is a flowchart of a method for a high-precision positioning road cone robot path planning system based on deep learning proposed in this invention;

[0065] Figure 2 This is a schematic diagram of the energy distribution modeling of the spatiotemporal energy constraint field of a high-precision positioning cone robot path planning system based on deep learning proposed in this invention.

[0066] Figure 3 This is a flowchart of the energy-optimized path planning process for a high-precision positioning road cone robot path planning system based on deep learning, as proposed in this invention. Detailed Implementation

[0067] 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.

[0068] refer to Figures 1-3 A high-precision localization path planning system for traffic cone robots based on deep learning, comprising:

[0069] The visual acquisition and preprocessing module is used to acquire visual image data, GNSS positioning data, RTK differential data and IMU inertial measurement data of road scenes, and perform preprocessing.

[0070] The visual semantic recognition module is used to input preprocessed visual image data into the visual recognition network, perform lane area segmentation, obstacle detection, road edge extraction and semantic element recognition on the road scene, and generate a set of scene semantic information.

[0071] The multi-source positioning fusion module is used to perform multi-source positioning fusion based on preprocessed GNSS positioning data, RTK differential data and IMU inertial measurement data to obtain a sequence of robot positioning results.

[0072] The semantic map construction module is used to perform spatiotemporal matching of scene semantic information sets with localization result sequences to generate road semantic maps;

[0073] The spatiotemporal energy constraint construction module is used to construct a spatiotemporal energy constraint field based on the road semantic map, mapping the lane centerline, obstacle boundary and road cone placement points as energy attraction source, repulsion source and target energy attraction source respectively, forming a spatiotemporal energy distribution structure that can be continuously optimized;

[0074] The energy optimization path planning module is used to input the spatiotemporal energy constraint field into the energy optimization network and generate path planning results based on the principle of energy minimization.

[0075] The dynamic path correction module is used to correct the path planning results based on the real-time updated set of scene semantic information and the sequence of positioning results, and generate dynamic path planning results.

[0076] The trajectory tracking control module is used to input the dynamic path planning results into the trajectory tracking controller, generate a continuous sequence of control commands, and drive the road cone robot to run along the dynamic path.

[0077] In this embodiment, the visual acquisition and preprocessing module includes:

[0078] The system collects visual image data, GNSS positioning data, RTK differential data, and IMU inertial measurement data of the road scene. The visual image data is continuously collected at a preset frame rate, the GNSS positioning data is real-time geographic coordinates, and the IMU inertial measurement data is triaxial acceleration and angular velocity.

[0079] Perform standardization operations on visual image data;

[0080] Formatting operations are performed on GNSS positioning data, RTK differential data, and IMU inertial measurement data to unify data structure, sampling frequency, and coordinate format, remove abnormal data points, and ensure the integrity of each type of data in time and space dimensions.

[0081] Time synchronization is performed on GNSS positioning data, RTK differential data and IMU inertial measurement data, and the timestamps of each data source are unified to a global time reference to form a time-synchronized multi-source positioning data sequence;

[0082] Spatial coordinate registration is performed on the multi-source positioning data sequence after time synchronization to establish the external parameter calibration relationship between the GNSS positioning coordinate system, RTK differential coordinate system and IMU inertial coordinate system, and the coordinate systems are uniformly mapped to the global coordinate system to achieve unified data alignment.

[0083] In this embodiment, standardized operations include color equalization, brightness normalization, noise suppression, and lens distortion correction.

[0084] In this embodiment, the visual semantic recognition module includes:

[0085] The preprocessed visual image data is input into the visual recognition network to perform feature extraction and semantic encoding on the road scene, generating a visual feature map that includes spatial and semantic features;

[0086] Visual recognition networks are deep learning neural network structures that extract features and perform semantic classification on input images to identify and locate target objects and region attributes in a scene.

[0087] In the lane region segmentation stage, pixel-by-pixel semantic classification is performed based on visual feature mapping to identify and divide the lane center region, lane boundary region and non-lane region in the road scene, generating a lane region segmentation result that includes lane region labels and pixel-level classification results.

[0088] In the obstacle detection stage, visual feature mapping is used to generate candidate regions and identify targets, and vehicles, pedestrians, traffic cones and fixed obstacles in the road scene are identified and located, and the obstacle detection results containing target category labels and boundary positions are output.

[0089] In the road edge extraction stage, the road edge pixel sequence is extracted based on the gradient change and geometric continuity information in the visual feature mapping. The curve fitting method is used to generate a road edge feature set to describe the spatial morphology of the road boundary.

[0090] In the semantic element recognition stage, the lane area segmentation results, obstacle detection results, and road edge feature sets are integrated and fused according to spatial correspondence to identify semantic elements of lane areas, road edges, and obstacles, generating a set of scene semantic information containing semantic labels, pixel-level segmentation results, and spatial coordinates.

[0091] In this embodiment, the multi-source localization fusion module includes:

[0092] The robot's initial position in the geographic coordinate system is calculated based on the preprocessed GNSS positioning data, and the GNSS initial positioning result is generated. The GNSS positioning data is used to provide a global position reference.

[0093] The preprocessed RTK differential data is used to perform differential correction on the initial GNSS positioning results to obtain high-precision position coordinates;

[0094] High-precision position coordinates provide spatial position information with centimeter-level positioning accuracy;

[0095] The robot's attitude angles are calculated based on the preprocessed IMU inertial measurement data, resulting in attitude angles including roll angle, pitch angle and yaw angle;

[0096] Roll angle, pitch angle, and yaw angle represent the rotation angles of the robot around the front-rear axis, left-right axis, and vertical axis, respectively, and are used to describe the robot's attitude and orientation in three-dimensional space.

[0097] Multi-source positioning fusion is performed on high-precision position coordinates and attitude angles. GNSS positioning data, RTK differential data and IMU inertial measurement data are fused and solved to generate positioning results that include robot three-dimensional coordinates and attitude angles.

[0098] The location results are organized in chronological order to form a location result sequence.

[0099] In this embodiment, the semantic map construction module includes:

[0100] The scene semantic information set and the localization result sequence are matched according to timestamps. The scene semantic information set is projected onto the global coordinate system to establish the spatiotemporal correspondence between the scene semantic information and the robot pose, and a global semantic mapping data set is generated.

[0101] The global coordinate system is a common three-dimensional reference coordinate frame that enables data from different sources to be aligned and calculated under the same spatial reference.

[0102] In the global semantic mapping dataset, the spatial point set corresponding to the semantic labels of the lane area is structurally extracted to form lane structure information containing the lane centerline and lane boundary.

[0103] Clustering and spatial localization are performed on the spatial point set corresponding to the semantic tags of obstacles to determine the spatial position and boundary shape of the obstacles in the global coordinate system and generate obstacle position information;

[0104] Based on lane structure information and obstacle location information, drivable area extraction is performed. Spatial difference and boundary calculation are performed on the lane centerline point set, boundary point set and obstacle point set to determine the drivable area boundary and generate drivable area information.

[0105] The location information of road cones is determined from the information of drivable areas. The coordinate set of road cone locations is generated according to the preset spacing, orientation and safety distance conditions, with the center line of the lane as the reference.

[0106] By spatiotemporally fusing and structurally combining lane structure information, obstacle location information, drivable area information, and traffic cone placement information, a road semantic map containing lane structure, obstacle location, drivable area, and traffic cone placement information is generated.

[0107] In this embodiment, the spatiotemporal energy constraint construction module includes:

[0108] Based on the lane structure, obstacle location, drivable area and road cone layout information in the road semantic map, a spatiotemporal energy constraint field coordinate framework is established in the global coordinate system.

[0109] The generation of the spatiotemporal energy constraint field coordinate framework specifically includes: calling the spatial coordinate data of various semantic elements in the road semantic map, and uniformly transforming the coordinates of lane centerlines, obstacle boundaries, drivable area boundaries, and road cone placement points to the global coordinate system; defining coordinate indexes for the spatial and temporal dimensions of the energy field based on the unified global coordinate system, and spatially dispersing the lane area, obstacle area, road boundary, and road cone placement points into a grid, generating a spatial raster structure containing fixed grid unit indices; assigning timestamp indices to each spatial grid unit according to time series information, forming a spatiotemporal joint coordinate index system; and establishing a spatiotemporal energy constraint field coordinate framework containing spatial location indexes, time indexes, and semantic element category indexes under the global coordinate system.

[0110] Weight parameters are defined according to the semantic element categories. The semantics of the lane area are mapped to the attraction weight parameters, the semantics of the obstacle are mapped to the repulsion weight parameters, and the semantics of the traffic cone placement points are mapped to the target attraction weight parameters, forming a set of semantic energy mapping parameters.

[0111] The attraction weight parameter, repulsion weight parameter, and target attraction weight parameter are set according to the importance ratio of the lane centerline, obstacle area, and road cone placement points in the task, and the sum of the three weight parameters is 1. The lane centerline and road cone placement points are both used as attraction energy sources. The former is used to form a continuous global attraction zone within the road area to guide the robot to move smoothly along the lane centerline, while the latter is used to form a local attraction extreme point in the target area to achieve precise convergence at the end of the path. The attraction weight parameter is used to control the global influence intensity of the attraction energy of the lane centerline, the target attraction weight parameter is used to enhance the local attraction effect at the road cone placement point, and the repulsion weight parameter is used to suppress energy interference in the obstacle area. The three types of weight parameters ensure that the path guidance and obstacle avoidance behavior remain balanced and stable in the energy constraint field.

[0112] The generation of the semantic energy mapping parameter set specifically includes: calling the semantic element category index in the spatiotemporal energy constraint field coordinate framework to identify lane area semantics, obstacle semantics, and cone placement point semantics respectively; assigning a corresponding energy weight type to each semantic element category in the global coordinate system, where lane area semantics corresponds to attraction weight parameters, used to generate attraction energy distribution to guide the robot to move along the center of the lane, obstacle semantics corresponds to repulsion weight parameters, used to generate repulsion energy distribution to restrict the robot from approaching the obstacle, and cone placement point semantics corresponds to target attraction weight parameters, used to generate local attraction centers to guide the robot to approach the placement target position; binding the above three types of weight parameters with their corresponding semantic space coordinates to construct an energy mapping table containing three weight parameters: attraction, repulsion, and target attraction; and importing the energy mapping table into the spatiotemporal energy constraint field coordinate framework to form a semantic energy mapping parameter set containing semantic category index, spatial location index, and energy weight parameter index.

[0113] Based on the lane structure information, the geometric curvature and directional continuity of the lane centerline direction are calculated, and the attraction energy intensity is allocated to the lane center region to form a continuous energy attraction distribution band in the lane center direction, so that the energy maintains a smooth gradient change in the center region, which constitutes the main attraction path of the robot in the spatiotemporal energy constraint field.

[0114] The generation of the main attraction path specifically includes: extracting the coordinate sequences of the lane centerline and the lane boundary lines; calculating the curvature change and direction change based on the spatial difference between adjacent centerline coordinate points; calculating the direction continuity parameter through the angle difference between two adjacent direction vectors; normalizing the curvature change and direction continuity parameters; weighting the normalized curvature change and direction continuity parameters to obtain a smoothness score; and generating an attraction energy intensity value by linear mapping within the interval [0,1] based on the magnitude of the smoothness score; using the lane centerline as the reference path, at each centerline sampling point, according to the corresponding attraction... The energy intensity value is linearly diffused along the normal direction, decreasing proportionally with distance to create a monotonically decaying distribution of the attracted energy within the range from the centerline to the boundary line. The attracted energy intensity of adjacent centerline sampling points is smoothed to ensure that the energy distribution is continuous in the spatial coordinate dimension and that the numerical changes are smooth. The spatial coordinates of the centerline sampling points and their attracted energy intensities are mapped to each other in the global coordinate system to generate a continuous attracted energy distribution band with the lane center direction as the main axis. This band forms the main attracted path of the robot in the spatiotemporal energy constraint field, which is used to guide the robot in path planning and motion control along the lane center direction.

[0115] Based on the obstacle's location information and geometric parameters, a local repulsion energy region is generated, and a repulsion energy rising zone is constructed around the obstacle, so that the energy gradually increases in the region close to the obstacle, forming a spatial repulsion zone;

[0116] The generation of the spatial repulsion zone specifically includes: extracting the center coordinates, boundary coordinates, and external dimensions of each obstacle; calculating the minimum radius from the obstacle center to the boundary as the initial repulsion radius; using the obstacle center as the origin, performing layered calculations on the outer side of the obstacle boundary in the global coordinate system according to a preset repulsion weight coefficient; at any point from the obstacle center, subtracting the initial repulsion radius from the Euclidean distance between the obstacle center and that point to obtain the distance difference; multiplying the distance difference by the repulsion weight coefficient to obtain the repulsion energy value, achieving a linear increase in energy as the distance difference decreases; and performing repulsion along the X and Y directions for rectangular obstacles. Energy superposition calculations are performed to average the radial angle of circular obstacles, ensuring a uniform numerical scale for the repulsion energy distribution of obstacles of different shapes. For situations where multiple obstacles coexist, the repulsion energy within overlapping areas is superimposed using a weighted method. The repulsion energy values ​​and spatial coordinates of each obstacle are mapped to the global coordinate system of the spatiotemporal energy constraint field, generating a repulsion energy rise band with continuous gradient variations. This enhances the energy value near the obstacle, forming a stable spatial repulsion zone. This zone is used to apply repulsion constraints to the robot's trajectory during path planning, guiding the robot to avoid obstacle areas and maintain a safe travel path.

[0117] The geometric boundaries of a rectangular obstacle are independently distributed in the X and Y directions. If the repulsive energy is calculated in only one direction, uneven energy distribution is likely to occur. It is necessary to calculate the repulsive energy distribution along the X and Y directions separately, and then add the results of the two directions together to allow the energy around the obstacle to diffuse simultaneously in the horizontal and vertical directions, so that the repulsive field forms a continuous "energy ring" around the rectangle, avoiding energy discontinuities or sharp gradients. A circular obstacle has radial symmetry in geometry. The repulsive energy is calculated along the radius with the center of the obstacle as the origin. However, in order to prevent the energy distribution in the angular direction from being uneven, it is necessary to take the energy values ​​in multiple directions at different angles and then average them at the angles. That is, the energy values ​​at each angle are averaged to keep the ring energy field changing continuously and smoothly around the circular obstacle.

[0118] Based on the location information of the road cones, a local energy attraction center is constructed in the global coordinate system, and a local energy minima are established at each road cone location, so that the energy forms a smoothly converging attraction potential trap at that location.

[0119] The establishment of the local minimum energy region specifically includes: extracting the coordinate set of all road cone placement points in the road semantic map; calculating the spatial position value of each placement point in the global coordinate system; setting an initial attraction energy intensity value at the center of each placement point, using this value as the initial intensity of the local energy attraction center; dividing the space into layers according to the Euclidean distance between the placement point and the surrounding space points; dividing the initial attraction energy intensity value by the distance layer number for each layer of space points to obtain the attraction energy attenuation value of each layer; performing a weighted average on the attraction energy attenuation value of each space point, and adding the weighted result to the initial attraction energy intensity value to form a smooth attraction energy distribution; performing minimum value filtering on the attraction energy distribution of the area surrounding each placement point, and taking the point with the minimum energy as the center of the local minimum attraction region; mapping the energy value and spatial coordinates of the local minimum attraction region to the global coordinate system of the spatiotemporal energy constraint field, so that each road cone placement point forms an attraction potential well with smooth energy convergence, which is used to guide the robot to gradually approach and accurately reach the corresponding road cone placement point along a continuous trajectory during the path planning process;

[0120] The initial attraction energy intensity value is set according to the importance level of the road cone placement point. The distance level number refers to the concentric level number formed by dividing the space around the road cone placement point with a fixed radius interval, which is used to indicate the distance level order between the spatial point and the placement point.

[0121] By integrating the energy distribution of lane attraction, the energy distribution of obstacle repulsion, and the energy distribution of road cone placement points, a continuous and differentiable spatiotemporal energy constraint field is established.

[0122] The generation of the spatiotemporal energy constraint field structure specifically includes: extracting the attraction energy intensity values ​​of each sampling point in the lane attraction energy band, the repulsion energy intensity values ​​of each spatial point in the obstacle repulsion energy zone, and the attraction energy intensity values ​​at the locations of road cones, and unifying them into a global coordinate system to form an energy sampling set; performing adaptive weighted fusion of the three types of energy distributions based on attraction weight parameters, repulsion weight parameters, and target attraction weight parameters; determining the energy action area to which each spatial sampling point belongs for each spatial sampling point in the global coordinate system; and calculating the attraction energy intensity and local repulsion energy when the sampling point is located in the lane attraction area. The energy values ​​are weighted and summed. When the sampling point is located in the area affected by the obstacle, the repulsive energy intensity is used as the main value, and the attractive energy intensity is subtracted to obtain the repulsion-dominant energy value in that area. When the sampling point is located in the neighborhood of the road cone placement point, the weighted average of the attractive energy intensity of the placement point and the attractive energy intensity of the lane is calculated to generate a locally convergent energy value. The energy values ​​of all spatial sampling points are normalized and smoothed to ensure continuous energy variation. A spatiotemporal energy constraint field containing the attraction-dominant region, the repulsion-dominant region, and the locally convergent region is formed in the global coordinate system, providing a stable energy constraint basis for path planning.

[0123] In the spatiotemporal energy-constrained field, the lane centerline, obstacle area, and cone placement points have different energy distribution characteristics. The lane centerline is given a high attraction energy intensity to form the dominant guiding path for robot movement. The obstacle area corresponds to a high repulsion energy region to prevent the path from approaching and achieve obstacle avoidance. The cone placement points are defined as target attraction centers, and their energy values ​​are set to local minima, so that the energy around the point forms a smooth and convergent energy potential well structure. This guides the robot to approach the point during the path energy minimization optimization process, achieving stable attraction and accurate arrival at the cone placement target.

[0124] The spatiotemporal energy constraint field is a unified data structure used to describe the relationship between motion constraints and task objectives of a road cone robot in a road environment. In the spatial dimension, it characterizes the energy distribution relationship between semantic elements such as roads, obstacles, and road cone placement points, and in the temporal dimension, it reflects the dynamic evolution process with environmental changes. The constraint field includes three core types of data: energy distribution feature values, used to characterize the attraction effect of the lane centerline, the repulsion effect of the obstacle area, and the target attraction effect of the road cone placement points; spatial position coordinates, used to calibrate the positional relationship of energy sampling points in the global coordinate system; and time index, used to record the change sequence of energy distribution at continuous time points, realizing the temporal modeling of the dynamic road environment. Through the unified expression of the three types of data, the spatiotemporal energy constraint field can simultaneously represent path guidance, obstacle avoidance constraints, and target attraction relationships in the same continuous space.

[0125] In this embodiment, the energy optimization path planning module includes:

[0126] The spatiotemporal energy constraint field is input into the energy optimization network, which performs path optimization calculations based on the energy distribution characteristic values, spatial location coordinates, and time index in the spatiotemporal energy constraint field.

[0127] Based on the energy distribution characteristics in the spatiotemporal energy constraint field, the energy difference between adjacent spatial coordinates is weighted and normalized to calculate the path energy cost, which is used to characterize the energy constraint strength of the path in the spatiotemporal energy constraint field.

[0128] The generation of path energy cost specifically includes: extracting the energy distribution characteristic value of each path node according to the spatial location coordinates of the path nodes based on the energy distribution characteristic value in the spatiotemporal energy constraint field; performing adjacent difference operation on the energy distribution characteristic values ​​of all path nodes to calculate the energy difference between nodes; weighting and summing the node energy differences according to the distance between nodes to obtain the weighted total energy value; and normalizing the weighted total energy value to generate the path energy cost value, which is used to characterize the overall energy consumption degree of the path in the spatiotemporal energy constraint field.

[0129] The path node sequence is iteratively updated based on the principle of minimizing energy until the absolute value of the difference between the path energy cost of two consecutive iterations is less than the preset convergence threshold, thus generating the optimal path node set.

[0130] The generation of the optimal path node set specifically includes: extracting the energy distribution feature values ​​of adjacent path nodes and calculating the energy difference between nodes; determining the energy gradient direction of each node based on the energy difference and calculating the path energy change to characterize the energy adjustment magnitude of the path in the current iteration; determining the node position correction step size based on the absolute value of the energy change and updating the spatial coordinates of the path nodes along the energy gradient descent direction to generate an updated path node sequence; performing a weighted average on the difference between the node coordinates before and after the update to obtain a smoothing correction amount, and superimposing the smoothing correction amount onto the updated node coordinates to form a smoothed correction path node sequence; repeating the energy difference calculation, energy gradient update, and node coordinate correction process until the absolute value of the path energy cost difference between two consecutive iterations is less than a preset convergence threshold, thus generating the optimal path node set;

[0131] Based on the spatial coordinates of the energy-optimal path node set, curve fitting is performed to calculate the trajectory curvature and orientation angle of each path node, and a path node velocity distribution sequence is generated based on the time index.

[0132] The path node sequence and velocity distribution sequence are combined to output the path planning result, which includes the path node sequence, trajectory curvature, and velocity distribution sequence.

[0133] In this embodiment, the dynamic path correction module includes:

[0134] Receives a set of scene semantic information and a sequence of positioning results that are updated in real time;

[0135] The real-time scene semantic information set and the localization result sequence are compared with the current path planning result to identify the affected areas in the path;

[0136] The lane structure, obstacle location, and traffic cone placement information are updated based on the latest scene semantic information set, and the positions of affected path nodes are adjusted.

[0137] Based on the adjusted path node sequence, the path direction and node connection order are recalculated to generate the corrected path planning result;

[0138] The corrected path planning results are matched and verified with the real-time positioning results to ensure the continuity and feasibility of the path, and a dynamic path planning result is generated. The dynamic path planning result refers to the real-time executable path scheme obtained by correcting the original path in real time based on the real-time updated scene semantic information and positioning results during the robot's movement.

[0139] To verify the feasibility of this invention, it was applied to a road safety deployment scenario in a city's intelligent traffic management center. The test area was selected on a main urban road section, characterized by typical urban road features such as multiple lanes, high traffic volume, and complex obstacle distribution. When performing its tasks, the road cone robot needs to complete precise deployment, dynamic obstacle avoidance, and path self-correction based on real-time road conditions. Traditional robot systems in this environment often rely on static maps and fixed paths. When lane drift, obstacle changes, or GNSS signal obstruction occur, it is easy to cause the accumulation of positioning errors and a decrease in deployment accuracy. However, this invention integrates GNSS positioning data, RTK differential data, IMU inertial measurement data, and visual image data to construct a multi-source synchronous input sequence, achieving data unification and complementarity in the spatiotemporal dimensions.

[0140] During implementation, the visual semantic recognition module of this invention can identify lane areas, obstacles, and traffic cone placement points in real time, generating a set of scene semantic information containing semantic labels and spatial coordinates. The multi-source positioning fusion module superimposes IMU attitude information based on the high-precision solution results of GNSS and RTK to form centimeter-level positioning results. The semantic map construction module performs spatiotemporal matching on these results to generate a road semantic map containing lane structure and obstacle distribution. Through spatiotemporal energy constraint field modeling, the system maps the lane centerline, obstacle area, and placement points as attraction sources, repulsion sources, and target attraction sources, respectively, realizing a unified expression of path search and physical constraints. The energy optimization network solves the path under the principle of energy minimization, generating a smooth, continuous, and differentiable driving trajectory. The dynamic path correction module adjusts the path using real-time updated semantic information and positioning data to maintain trajectory continuity and safety.

[0141] To verify the performance of the present invention in practice, it was compared with traditional methods.

[0142] Table 1. Performance Comparison of the Invention and Traditional Methods

[0143]

[0144] As can be seen from Table 1, the high-precision positioning road cone robot path planning system based on deep learning proposed in this invention outperforms traditional path planning methods in several core performance indicators.

[0145] In terms of path accuracy, the average path offset error of the system was reduced from 0.42m in the traditional method to 0.11m, and the deployment position accuracy was improved from 0.35m to 0.08m, indicating that the multi-source localization fusion mechanism that integrates GNSS, RTK and IMU data has improved the robot's perception accuracy of its own position.

[0146] In terms of path continuity and obstacle avoidance capability, the energy optimization network uses a spatiotemporal energy constraint field to solve the path, which improves the path smoothness by 22.4% and the obstacle avoidance success rate to 97.4%. Traditional methods rely only on local planning or static maps, which are prone to path changes or obstacle avoidance failures in complex traffic environments. In contrast, this invention guides the robot to achieve natural turning and smooth adjustment in a field of attraction and repulsion balance through energy distribution.

[0147] In terms of response speed and system robustness, the path planning response time of this invention is reduced from 1.82s to 0.95s, and the stability of dynamic scene replanning is improved from 82.3% to 96.8%. This is due to the rapid iterative update mechanism of the energy optimization network and its real-time processing capability of spatiotemporal semantic information, which enables the system to quickly adjust the path without losing control continuity in complex environments such as lane drift and the appearance of temporary obstacles.

[0148] Furthermore, the multi-source localization fusion error decreased to 0.07m, indicating that the coupling correction of RTK differential and IMU inertial calculation effectively reduced localization jitter, and the overall energy consumption decreased by 26.8%, which shows that path smoothness and planning rationality directly reduced the robot's ineffective energy consumption during movement.

[0149] The above are merely preferred embodiments 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 high-precision localization path planning system for a traffic cone robot based on deep learning, characterized in that, include: The visual acquisition and preprocessing module is used to acquire visual image data, GNSS positioning data, RTK differential data and IMU inertial measurement data of road scenes, and perform preprocessing. The visual semantic recognition module is used to input preprocessed visual image data into the visual recognition network, perform lane area segmentation, obstacle detection, road edge extraction and semantic element recognition on the road scene, and generate a set of scene semantic information. The multi-source positioning fusion module is used to perform multi-source positioning fusion based on preprocessed GNSS positioning data, RTK differential data and IMU inertial measurement data to obtain a sequence of robot positioning results. The semantic map construction module is used to perform spatiotemporal matching of scene semantic information sets with localization result sequences to generate road semantic maps; The spatiotemporal energy constraint construction module is used to construct a spatiotemporal energy constraint field based on the road semantic map, mapping the lane centerline, obstacle boundary and road cone placement points as energy attraction source, repulsion source and target energy attraction source respectively, forming a spatiotemporal energy distribution structure that can be continuously optimized; The energy optimization path planning module is used to input the spatiotemporal energy constraint field into the energy optimization network and generate path planning results based on the principle of energy minimization. The dynamic path correction module is used to correct the path planning results based on the real-time updated set of scene semantic information and the sequence of positioning results, and generate dynamic path planning results. The trajectory tracking control module is used to input the dynamic path planning results into the trajectory tracking controller, generate a continuous sequence of control commands, and drive the road cone robot to run along the dynamic path.

2. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The visual acquisition and preprocessing module includes: Collect visual image data, GNSS positioning data, RTK differential data, and IMU inertial measurement data of road scenes; Perform standardization operations on visual image data; Perform formatting operations on GNSS positioning data, RTK differential data, and IMU inertial measurement data; Time synchronization is performed on GNSS positioning data, RTK differential data and IMU inertial measurement data, and the timestamps of each data source are unified to a global time reference to form a time-synchronized multi-source positioning data sequence; Spatial coordinate registration is performed on the time-synchronized multi-source positioning data sequence.

3. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 2, characterized in that, The standardized operations include color equalization, brightness normalization, noise suppression, and lens distortion correction.

4. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The visual semantic recognition module includes: The preprocessed visual image data is input into the visual recognition network to perform feature extraction and semantic encoding on the road scene, generating a visual feature map that includes spatial and semantic features; In the lane region segmentation stage, pixel-by-pixel semantic classification is performed based on visual feature mapping to identify and divide the lane center region, lane boundary region and non-lane region in the road scene, generating a lane region segmentation result that includes lane region labels and pixel-level classification results. In the obstacle detection stage, visual feature mapping is used to generate candidate regions and identify targets, and vehicles, pedestrians, traffic cones and fixed obstacles in the road scene are identified and located, and the obstacle detection results containing target category labels and boundary positions are output. In the road edge extraction stage, the road edge pixel sequence is extracted based on the gradient change and geometric continuity information in the visual feature mapping, and the road edge feature set is generated by the curve fitting method. In the semantic element recognition stage, the lane area segmentation results, obstacle detection results, and road edge feature sets are integrated and fused according to spatial correspondence to identify semantic elements of lane areas, road edges, and obstacles, generating a set of scene semantic information containing semantic labels, pixel-level segmentation results, and spatial coordinates.

5. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The multi-source positioning fusion module includes: Calculate the robot’s initial position in the geographic coordinate system based on the preprocessed GNSS positioning data, and generate the GNSS initial positioning result. The preprocessed RTK differential data is used to perform differential correction on the initial GNSS positioning results to obtain high-precision position coordinates; The robot's attitude angles are calculated based on the preprocessed IMU inertial measurement data, resulting in attitude angles including roll angle, pitch angle and yaw angle; Multi-source positioning fusion is performed on high-precision position coordinates and attitude angles. GNSS positioning data, RTK differential data and IMU inertial measurement data are fused and solved to generate positioning results that include robot three-dimensional coordinates and attitude angles. The location results are organized in chronological order to form a location result sequence.

6. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The semantic map construction module includes: The scene semantic information set and the localization result sequence are matched according to timestamps. The scene semantic information set is projected onto the global coordinate system to establish the spatiotemporal correspondence between the scene semantic information and the robot pose, and a global semantic mapping data set is generated. In the global semantic mapping dataset, the spatial point set corresponding to the semantic labels of the lane area is structurally extracted to form lane structure information containing the lane centerline and lane boundary. Clustering and spatial localization are performed on the spatial point set corresponding to the semantic tags of obstacles to determine the spatial position and boundary shape of the obstacles in the global coordinate system and generate obstacle position information; Based on lane structure information and obstacle location information, drivable area extraction is performed. Spatial difference and boundary calculation are performed on the lane centerline point set, boundary point set and obstacle point set to determine the drivable area boundary and generate drivable area information. The location information of road cones is determined from the information of drivable areas. The coordinate set of road cone locations is generated according to the preset spacing, orientation and safety distance conditions, with the center line of the lane as the reference. By spatiotemporally fusing and structurally combining lane structure information, obstacle location information, drivable area information, and traffic cone placement information, a road semantic map containing lane structure, obstacle location, drivable area, and traffic cone placement information is generated.

7. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The spatiotemporal energy constraint construction module includes: Based on the lane structure, obstacle location, drivable area and road cone layout information in the road semantic map, a spatiotemporal energy constraint field coordinate framework is established in the global coordinate system. Weight parameters are defined according to the semantic element categories. The semantics of the lane area are mapped to the attraction weight parameters, the semantics of the obstacle are mapped to the repulsion weight parameters, and the semantics of the traffic cone placement points are mapped to the target attraction weight parameters, forming a set of semantic energy mapping parameters. Based on the lane structure information, the geometric curvature and directional continuity of the lane centerline direction are calculated, and the attraction energy intensity is allocated to the lane center region to form a continuous energy attraction distribution band in the lane center direction, so that the energy maintains a smooth gradient change in the center region, which constitutes the main attraction path of the robot in the spatiotemporal energy constraint field. Based on the obstacle's location information and geometric parameters, a local repulsion energy region is generated, and a repulsion energy rising zone is constructed around the obstacle, so that the energy gradually increases in the region close to the obstacle, forming a spatial repulsion zone; Based on the location information of the road cones, a local energy attraction center is constructed in the global coordinate system, and a local energy minima are established at each road cone location, so that the energy forms a smoothly converging attraction potential trap at that location. By integrating the energy distribution of lane attraction, the energy distribution of obstacle repulsion, and the energy distribution of traffic cone placement points, a continuous and differentiable spatiotemporal energy constraint field is established.

8. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The energy optimization path planning module includes: The spatiotemporal energy constraint field is input into the energy optimization network, which performs path optimization calculations based on the energy distribution characteristic values, spatial location coordinates, and time index in the spatiotemporal energy constraint field. Based on the energy distribution characteristics in the spatiotemporal energy constraint field, the energy difference between adjacent spatial coordinates is weighted and normalized to calculate the path energy cost. The path node sequence is iteratively updated based on the principle of minimizing energy until the absolute value of the difference between the path energy cost of two consecutive iterations is less than the preset convergence threshold, thus generating the optimal path node set. Based on the spatial coordinates of the energy-optimal path node set, curve fitting is performed to calculate the trajectory curvature and orientation angle of each path node, and a path node velocity distribution sequence is generated based on the time index. The path node sequence and velocity distribution sequence are combined to output the path planning result, which includes the path node sequence, trajectory curvature, and velocity distribution sequence.

9. The high-precision positioning path planning system for a traffic cone robot based on deep learning according to claim 1, characterized in that, The dynamic path correction module includes: Receives a set of scene semantic information and a sequence of positioning results that are updated in real time; The real-time scene semantic information set and the localization result sequence are compared with the current path planning result to identify the affected areas in the path; The lane structure, obstacle location, and traffic cone placement information are updated based on the latest scene semantic information set, and the positions of affected path nodes are adjusted. Based on the adjusted path node sequence, the path direction and node connection order are recalculated to generate the corrected path planning result; The revised path planning results are matched and verified with the real-time positioning results to ensure path continuity and feasibility, and dynamic path planning results are generated.

Images