Intelligent parking space navigation system and method for new energy commercial vehicle

Abstract

The invention discloses a parking space intelligent navigation system and method for a new energy commercial vehicle, and relates to the technical field of intelligent parking. Projecting the point cloud data acquired by any radar sensor to a two-dimensional plane to obtain a projection image set; acquiring images acquired by each visual sensor to obtain a visual image set; fusing the projection image and the visual image to obtain a fused feature map; inputting the fused feature map into a target detection model to obtain a feature tag, and adding the feature tag into the fused feature map to obtain a target feature map; constructing according to the target feature image set to obtain a parking environment and determining a target constraint condition; and performing parking space navigation optimization according to the constraint condition to obtain a target path, and performing parking space navigation according to the target path. The point cloud data is projected to the two-dimensional plane to be fused with the visual image, the efficiency and accuracy of the detection model are improved through the input of the fused feature map, more accurate constraints are provided for the optimization of the parking space path, the planning time is shortened, and the success rate of the parking path is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent parking technology, specifically to an intelligent parking space navigation system and method for new energy commercial vehicles. Background Technology

[0002] With the rapid development of intelligent transportation systems, automated parking technology has gradually become an important research direction in the field of autonomous driving. From early simple assisted parking based on ultrasonic sensors to today's complex environmental perception and path planning that integrates multiple sensors, automated parking technology has evolved from a single function to integration and intelligence.

[0003] Patent No. CN118061988A discloses a parking path optimization method; Step 1: Record scene information and establish a parking map and a corner point map of parking space entrances; Step 2: Obtain the lane center line, left parking space line, and right parking space line from the scene information, and generate reference line information; Step 3: Determine the available reference curves on the left and right sides based on the reference line information; Step 4: Select the corresponding available reference line according to the parking type; Step 5: Select the corresponding optimized path based on the available reference line and lane type.

[0004] In existing technologies, during the parking process of new energy commercial vehicles, the point cloud data scanned by lidar and the visual images collected by visual sensors cannot be directly fused, resulting in low efficiency and reduced accuracy of the detection model, which leads to poor subsequent path optimization and long planning time. Summary of the Invention

[0005] The purpose of this invention is to solve the problem mentioned in the background art that the point cloud data scanned by lidar and the visual images collected by visual sensors cannot be directly fused, resulting in low efficiency and reduced accuracy of the detection model, which leads to poor effect of subsequent path optimization and long planning time. Therefore, this invention proposes a parking space intelligent navigation system and method for new energy commercial vehicles.

[0006] A first aspect of this invention provides a parking space intelligent navigation method for new energy commercial vehicles. The new energy commercial vehicle is equipped with multiple sensor groups, each sensor group including a visual sensor and a radar sensor, with the data acquisition angles of the visual sensor and the radar sensor corresponding one-to-one. The method includes: Point cloud data collected by any radar sensor is projected onto a two-dimensional plane to obtain a projected image set; A visual image set is obtained by acquiring images from each visual sensor; The projected image and visual image corresponding to each sensor group are fused to obtain a fused feature map; The fused feature map is input into the target detection model to obtain feature labels, and the feature labels are added to the fused feature map to obtain the target feature map; A target feature map set is determined, a parking environment is constructed based on the target feature map set, and target constraints are determined based on the parking environment; the target feature map set includes target feature maps from any angle collected by new energy commercial vehicles; The target path is obtained by optimizing parking space navigation based on the constraints, and parking space navigation is performed based on the target path.

[0007] Optionally, projecting the point cloud data onto a two-dimensional plane to obtain a projected image includes: Determine the target spherical coordinates of the target point cloud in the point cloud data, and transform the target spherical coordinates to obtain the target planar coordinates; the target point cloud is any one of the points in the point cloud data; An initial projected image is obtained by stitching together the coordinates of each plane, and the initial projected image is then transformed to obtain the final projected image.

[0008] Optionally, the projected image and the visual image are fused to obtain a fused feature map, including: The projected image is subjected to target operations a preset number of times to obtain a projected feature map, and the visual image is input into the patch embedding module to obtain a first feature map. The first feature map is then input into the target attention module to obtain a second feature map. The second feature map is sequentially input into the patch fusion module and the target attention module to obtain the third feature map. The third feature map is sequentially input into the patch fusion module and the target attention module to obtain the fourth feature map. The fourth feature map is sequentially input into the patch fusion module and the target attention module to obtain the fifth feature map. The projected feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map are input into the context aggregation module to obtain the fused feature map.

[0009] Optionally, the target constraints are determined based on the parking environment, including: The physical properties of the new energy commercial vehicle are obtained, basic constraints are determined based on the physical properties, and boundary constraints are determined based on the parking environment; the physical properties include: minimum turning radius, maximum curvature, and vehicle planar profile; The target constraints are determined based on the basic constraints and the boundary constraints.

[0010] Optionally, the target path is obtained by optimizing parking space navigation based on the constraints, including: The parameter combination is determined according to the constraints, and the parameter combination is used as a particle to generate an initial population according to the target constraints; the initial population contains multiple particles. The fitness of the target particles in the initial population is evaluated to obtain the fitness, and the particle with the lowest fitness is taken as the optimal particle; the target particle can be any one of the particles in the initial population. The initial population is iteratively updated according to the first preset number of iterations to obtain the judgment population, and the set of suboptimal particles in each generation is determined according to the first number of iterations; the set of suboptimal particles in each generation contains multiple suboptimal particles, and the suboptimal particle is the best individual of any particle in the iteration process. An update operation is performed on each suboptimal particle in the previous suboptimal particle set to obtain a target suboptimal particle set. The target optimal particle in the judgment population is determined, and the target optimal particle and the target suboptimal particle set are combined to obtain the target population. The target population is iteratively updated until a second preset number of iterations is reached, or if the change in fitness between the current generation's best particle and the previous generation's best particle is less than a preset threshold, then the optimal parameter combination is output, and the target path is determined based on the optimal parameter combination.

[0011] A second aspect of this invention provides a parking space intelligent navigation system for new energy commercial vehicles, the system comprising: The data projection module is used to acquire point cloud data of the target area and project the point cloud data onto a two-dimensional plane to obtain a projected image. The image fusion module is used to acquire a visual image of the target area and fuse the projected image and the visual image to obtain a fused feature map; the projected image and the visual image are in one-to-one correspondence in the acquisition direction; The labeling module is used to input the fused feature map into the target detection model to obtain feature labels, and to add the feature labels to the fused feature map to obtain the target feature map; The condition construction module is used to determine the target feature map set, construct a parking environment based on the target feature map set, and determine the target constraint conditions based on the parking environment; the target feature map set includes target feature maps from any angle collected by new energy commercial vehicles; The path optimization module is used to optimize parking space navigation according to the constraints to obtain the target path, and then perform parking space navigation according to the target path.

[0012] Optionally, the data projection module includes: The coordinate transformation module is used to determine the target spherical coordinates of the target point cloud in the point cloud data, and transform the target spherical coordinates to obtain the target planar coordinates; the target point cloud is any one of the points in the point cloud data; The image stitching module is used to stitch together an initial projected image based on the coordinates of each plane, and to perform coordinate transformation on the initial projected image to obtain a final projected image.

[0013] Optionally, the feature map fusion process includes: The projected image is subjected to target operations a preset number of times to obtain a projected feature map, and the visual image is input into the patch embedding module to obtain a first feature map. The first feature map is then input into the target attention module to obtain a second feature map. The second feature map is sequentially input into the patch fusion module and the target attention module to obtain the third feature map. The third feature map is sequentially input into the patch fusion module and the target attention module to obtain the fourth feature map. The fourth feature map is sequentially input into the patch fusion module and the target attention module to obtain the fifth feature map. The projected feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map are input into the context aggregation module to obtain the fused feature map.

[0014] Optionally, the condition construction module includes: The initial condition determination module is used to acquire the physical properties of the new energy commercial vehicle, determine the basic constraints based on the physical properties, and determine the boundary constraints based on the parking environment; the physical properties include: minimum turning radius, maximum curvature, and vehicle planar profile; The final condition determination module is used to determine the target constraint condition based on the basic constraint condition and the boundary constraint condition.

[0015] Optionally, the path optimization module includes: The first execution module is used to determine the parameter combination according to the constraint conditions, use the parameter combination as particles, and generate an initial population according to the target constraint conditions; the initial population contains multiple particles. The second execution module is used to evaluate the fitness of the target particles in the initial population and select the particle with the lowest fitness as the optimal particle; the target particle is any one of the particles in the initial population. The third execution module is used to iteratively update the initial population according to a first preset number of iterations to obtain a judgment population, and to determine the set of suboptimal particles in each generation according to the first number of iterations; the set of suboptimal particles in each generation contains multiple suboptimal particles, and the suboptimal particle is the best individual of any particle in the iteration process. The fourth execution module is used to perform update operations on each suboptimal particle in the previous suboptimal particle set to obtain the target suboptimal particle set, determine the target optimal particle in the judgment population, and combine the target optimal particle and the target suboptimal particle set to obtain the target population. The fifth execution module is used to iteratively update the target population until a second preset number of iterations is reached, or if the change in fitness between the current generation's best particle and the previous generation's best particle is less than a preset threshold, then the optimal parameter combination is output, and the target path is determined based on the optimal parameter combination.

[0016] The beneficial effects of this invention are: This invention proposes an intelligent parking space navigation method for new energy commercial vehicles. The method involves projecting point cloud data collected by any radar sensor onto a two-dimensional plane to obtain a projected image set; acquiring images collected by each visual sensor to obtain a visual image set; fusing the projected images and visual images to obtain a fused feature map; inputting the fused feature map into a target detection model to obtain feature labels, which are then added to the fused feature map to obtain a target feature map; constructing a parking environment and determining target constraints based on the target feature map set; optimizing parking space navigation based on the constraints to obtain a target path; and performing parking space navigation based on the target path. Projecting point cloud data onto a two-dimensional plane and fusing it with visual images, along with inputting the fused feature map, improves the efficiency and accuracy of the detection model, provides more precise constraints for parking space path optimization, shortens planning time, and increases the success rate of parking paths. Attached Figure Description

[0017] Figure 1 A flowchart of a parking space intelligent navigation method for new energy commercial vehicles is provided as an embodiment of the present invention; Figure 2 This invention provides a schematic diagram of a parking space intelligent navigation system for new energy commercial vehicles. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention provides a method for intelligent parking space navigation for new energy commercial vehicles. See also... Figure 1 , Figure 1 A flowchart illustrating an intelligent parking space navigation method for new energy commercial vehicles, provided as an embodiment of the present invention. The method includes the following steps: S101, Project the point cloud data collected by any radar sensor onto a two-dimensional plane to obtain a projected image set; S102, acquire the images collected by each vision sensor to obtain a visual image set; S103, fuse the projected image and visual image corresponding to each sensor group to obtain a fused feature map; S104, Input the fused feature map into the target detection model to obtain feature labels, and add the feature labels to the fused feature map to obtain the target feature map; S105, Determine the target feature map set, construct the parking environment based on the target feature map set, and determine the target constraints based on the parking environment; S106, optimize parking space navigation according to constraints to obtain the target path, and perform parking space navigation according to the target path.

[0020] The target feature map set contains target feature maps from any angle collected from new energy commercial vehicles; The present invention provides an intelligent parking space navigation method for new energy commercial vehicles. By projecting point cloud data onto a two-dimensional plane and fusing it with visual images, the input of the fused feature map improves the efficiency and accuracy of the detection model, provides more precise constraints for parking space path optimization, shortens planning time, and increases the success rate of parking paths.

[0021] In one implementation, point cloud data is acquired by radar sensors installed on the new energy commercial vehicle, and visual images are acquired by visual sensors. The fused feature map is input into a target detection model (e.g., a YOLO series detection model) for obstacle detection. The identified obstacles are added as feature labels to the fused feature map to obtain a target feature map. A three-dimensional scene is constructed based on the target feature maps collected from various angles by the new energy commercial vehicle. The three-dimensional scene is used as the parking environment. The constraints of the path optimization algorithm are determined based on the parking environment, and the parking space navigation path is optimized accordingly.

[0022] In one implementation, the target area is the region where data is collected from the new energy commercial vehicle, such as the front, left, right, and rear sides of the vehicle. A projected image is obtained by projecting the point cloud data of the target area onto a two-dimensional plane. This image fusion, combined with a visual image, achieves multimodal data fusion. This image fusion method not only retains the spatial perception advantages of point cloud data but also integrates the richness of texture and semantic information in visual images. This allows the target detection model to indirectly recognize point cloud data, while the visual image provides visual texture and semantic information. The fused feature map can more comprehensively reflect the environmental information of the target area, providing richer and more accurate input data for subsequent target detection and path planning, thus improving the system's adaptability and robustness.

[0023] In one implementation, inputting the fused feature map into the target detection model enables more accurate identification of key elements within the target area, such as vehicles and obstacles. By adding feature labels to the fused feature map, a target feature map is generated, enhancing the model's ability to identify targets. The feature map formed by projecting radar point cloud data is fused with the feature map of the visual image, allowing the target detection model to simultaneously utilize the ranging accuracy of radar and the semantic information of the image. This helps improve the accuracy of subsequent detection processes and also enhances the target recognition capabilities of the target detection model in complex environments, providing data support for the construction of parking environment models.

[0024] In one implementation, a parking environment is constructed based on a target feature map set, which can comprehensively reflect the characteristics of the new energy commercial vehicle and its surrounding environment from multiple perspectives. The target feature map set contains the target feature map of the new energy commercial vehicle at any angle. The aforementioned feature map covers the feature information of the vehicle in different positions and postures, thereby dynamically reflecting environmental changes.

[0025] In one implementation, the parking environment model constructed in the above manner can more accurately describe the spatial relationship between the vehicle and surrounding obstacles, parking space boundaries, and other elements, providing more accurate and complete environmental information for path planning, enhancing the system's adaptability and flexibility, and enabling the parking environment model to maintain high accuracy and robustness even in complex scenarios.

[0026] In one implementation, parking space navigation optimization is performed based on constraints determined by the parking environment model, which can generate a more reasonable and efficient target path. The feature labels fused in the feature map provide clear constraints for path planning, including information such as obstacle boundaries, drivable areas, and relative speeds. Based on these constraints, the path planning algorithm can directly call upon the above information in a unified two-dimensional feature space without the need for additional collision detection or speed filtering, thereby shortening the pathfinding time. By optimizing path planning, vehicles can quickly generate parking paths in complex parking environments, improving parking efficiency and success rate.

[0027] In one embodiment, projecting point cloud data onto a two-dimensional plane to obtain a projected image includes: Determine the target spherical coordinates of the target point cloud in the point cloud data, and transform the target spherical coordinates to obtain the target planar coordinates; the target point cloud can be any one of the points in the point cloud data. The initial projected image is obtained by stitching together the coordinates of each plane, and the coordinate transformation is performed on the initial projected image to obtain the final projected image.

[0028] In one implementation, the target spherical coordinates of the target point cloud in the point cloud data are determined. Specifically, the target spherical coordinates are the three-dimensional spatial parameters detected by the radar. That is, the spherical coordinates of each target point cloud are represented as (r, θ, Φ), where r is the radial distance between the radar and the target point, θ is the horizontal azimuth angle (the angle between the target point in the horizontal direction and the radar's central axis), and Φ is the vertical elevation angle (the angle between the target point in the vertical direction and the radar's horizontal plane). The target spherical coordinates are transformed to obtain the target planar coordinates. Specifically, the spherical coordinates are converted to 2D polar coordinates through circular projection and then mapped to Cartesian coordinates: the elevation angle Φ is ignored (based on the radar's weak altitude perception), r and θ are retained, and (r, θ) is first used as polar coordinate parameters. Then, the two-dimensional planar Cartesian coordinates (x, y) are obtained through the calculation of (x=r*cosθ, y=r*sinθ).

[0029] In one implementation, an initial projected image is obtained by stitching together the coordinates of each plane. The stitching process involves expanding the plane coordinates (x, y) of a single target point cloud into a small circular region, increasing the projection density based on a preset radius, and filtering out noise points with radii smaller than a threshold to avoid interference from false information. The initial projected image is then transformed to obtain the projected image, specifically through coordinate system alignment transformation: the initial projected image in the radar coordinate system is transformed to the global coordinate system using rotation and translation matrices, and then transformed to the camera image coordinate system using inverse rotation, inverse translation, and camera intrinsic parameter matrices (including parameters such as focal length and principal point), ultimately obtaining a projected image aligned with the camera image space.

[0030] In one implementation, a structured transformation from radar 3D point cloud to 2D image is achieved through precise conversion and stitching optimization of "spherical coordinates to planar coordinates," solving the dimensional compatibility problem between radar point cloud and visual image. Through coordinate system alignment, spatial consistency between radar projection image and camera image is ensured, improving feature matching accuracy during fusion. For example, the projection position of an obstacle 5 meters ahead detected by radar is aligned with the pixel position of the same obstacle captured by camera, reducing positional deviation during subsequent feature fusion and avoiding target mismatch due to coordinate misalignment.

[0031] In one embodiment, fusing the projected image and the visual image to obtain a fused feature map includes: The projected image is subjected to target operations a preset number of times to obtain a projected feature map. The visual image is then input into the patch embedding module to obtain a first feature map. The first feature map is then input into the target attention module to obtain a second feature map. The second feature map is sequentially input into the patch fusion module and the target attention module to obtain the third feature map. The third feature map is sequentially input into the patch fusion module and the target attention module to obtain the fourth feature map. The fourth feature map is sequentially input into the patch fusion module and the target attention module to obtain the fifth feature map. The projected feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map are input into the context aggregation module to obtain the fused feature map.

[0032] In one implementation, the projected image is subjected to target operations a preset number of times to obtain a projected feature map. The preset number of times can be five or six times, etc. The target operations are: convolution (Conv), batch normalization (Batch Normalization), and ReLU activation. After the visual image is initially feature-encoded by the patch embedding module, the feature map is processed multiple times by the patch merging module and the target attention module (BuildFormer block). From the second feature map to the fifth feature map, the patch merging module gradually reduces the spatial resolution of the feature map and aggregates local features. At the same time, the attention mechanism of the target attention module captures global correlations at different scales. Multi-scale features from high-resolution local details to low-resolution global structures are extracted from the visual image, so that the second to fifth feature maps correspond to different levels of visual information, providing a richer feature foundation for subsequent fusion.

[0033] In one implementation, the projection feature map originates from radar point cloud projection and contains core radar information such as distance and velocity. The projection feature map and the second to fifth feature maps, which originate from visual images and contain visual information such as texture and semantics, are fused by the input context aggregation module. The radar features can provide stable target position information, while the visual features can provide target details and semantics. The multi-scale characteristics of the second to fifth feature maps further cover different levels of visual information. By fusing, the advantages of the two sensors can be complemented, reducing noise from a single sensor, such as visual noise in low-light environments and interference caused by radar sparsity, thereby improving the robustness of perception in complex environments.

[0034] In one implementation, the patch fusion module reduces the feature map resolution by merging small regions of the image, thereby reducing the computational load of the subsequent target attention module and avoiding redundant processing of pixels. Meanwhile, the target attention module focuses on key regions through a window-based attention mechanism, preserving core features while reducing computational complexity. The iterative process of "patch fusion module + target attention module" gradually improves the feature abstraction level while reducing computational overhead through structural optimization. Finally, the context aggregation module processes multiple feature maps, avoiding the inefficiency of directly fusing the original data.

[0035] In one embodiment, determining the target constraints based on the parking environment includes: The physical properties of the new energy commercial vehicle are obtained, the basic constraints are determined based on the physical properties, and the boundary constraints are determined based on the parking environment. The physical properties include: minimum turning radius, maximum curvature, and vehicle planar profile. Determine the target constraints based on the basic constraints and boundary constraints.

[0036] In one implementation, the physical properties of new energy commercial vehicles are as follows: the minimum turning radius is determined by the vehicle's wheelbase and maximum steering angle; in a two-dimensional plane, the radius of the arc segment must be ≥ R. min Otherwise, the vehicle cannot complete the steering; maximum curvature (ρ max ): Curvature ρ=1 / R, therefore ρ max =1 / R min The curvature of any point on the path cannot exceed this value to prevent oversteering; Vehicle planar profile (length L, width W): With the center of the rear axle of the vehicle as the origin, the planar coordinates of the front end, rear end, left and right ends can be determined by geometric calculation. All points on the path must ensure that the vehicle profile does not exceed its own size limit to avoid the "self-collision" logic error.

[0037] One implementation ensures the physical feasibility and environmental adaptability of the parking path. By using basic constraints determined by the vehicle's physical attributes (such as minimum turning radius and maximum curvature), the problem of the path exceeding the vehicle's movement capabilities (such as impossible steering maneuvers) is fundamentally avoided, ensuring the path's executability at the vehicle's physical level. Combined with boundary constraints determined by the parking environment, the risk of collision between the path and environmental obstacles (such as parking space boundaries and adjacent vehicles) is specifically avoided.

[0038] One implementation improves the efficiency and accuracy of parking path planning. Layered determination of basic and boundary constraints avoids computational redundancy caused by treating all constraints the same. Basic constraints fix the core physical limitations of the vehicle (e.g., the minimum turning radius does not change with environmental variations), while boundary constraints focus on dynamic environmental constraints (e.g., the boundary differences between different parking spaces). The integrated objective constraints clearly define the effective search range of path parameters. This eliminates the need for subsequent path optimization algorithms (such as hybrid PSO-GA) to search within invalid ranges, reducing unnecessary computational costs. Furthermore, the precise constraint range allows for faster convergence to the optimal path, improving overall planning efficiency.

[0039] In one embodiment, optimizing parking space navigation based on constraints to obtain a target path includes: The parameter combination is determined according to the constraints, and the parameter combination is used as a particle to generate an initial population according to the target constraints; the initial population contains multiple particles. The fitness of the target particles in the initial population is evaluated, and the particle with the lowest fitness is selected as the optimal particle; the target particle can be any one of the particles in the initial population. The initial population is iteratively updated according to the first preset number of iterations to obtain the judgment population. The set of suboptimal particles in each generation is determined according to the first number of iterations. The set of suboptimal particles in each generation contains multiple suboptimal particles, and the suboptimal particle is the best individual of any particle in the iteration process. The target suboptimal particle set is obtained by performing an update operation on each suboptimal particle in the previous suboptimal particle set, determining the target optimal particle in the judgment population, and combining the target optimal particle and the target suboptimal particle set to obtain the target population. The target population is iteratively updated until the second preset number of iterations is reached, or if the change in fitness between the current generation's best particle and the previous generation's best particle is less than a preset threshold, then the optimal parameter combination is output, and the target path is determined based on the optimal parameter combination.

[0040] In one implementation, the parameter combination is determined according to the constraints: line length: the straight-line distance of the vehicle; arc segment radius: the radius of curvature of the turning arc; arc segment length: the arc length of the turning arc. The above parameter combination directly determines a complete parking path, so optimizing it can yield the optimal path.

[0041] In one implementation, for example: the particle is a three-dimensional vector (line length, arc radius, arc length), with each dimension corresponding to an optimization parameter; 30 particles are randomly generated based on the constraint interval (e.g., L in [L... min ,L max R takes a random value within [R], and R is within [R]. min ,R max The parameters are randomly selected within the range to ensure that all particles' parameters meet the constraints (to avoid infeasible initial parameters); for any particle, the velocity needs to be initialized, which is the adjustment range of the corresponding parameter and is used to control the step size of parameter updates.

[0042] In one implementation, a corresponding parking path is generated for each particle (parameter combination), and the quality of the path is evaluated by a fitness function, with a smaller value indicating better performance. The fitness function is: F = α*C1 + β*C2 + γ*C3, where F is the fitness, α, β, and γ are constant proportionality coefficients, C1 is the path smoothness, C2 is the economic cost, and C3 is the space utilization cost. The smoothness cost is calculated by the maximum value and average rate of change of the path curvature; the more continuous the curvature and the smaller the change, the smaller C1. The economic cost is calculated by the total path length; the shorter the length, the smaller C2. The space utilization cost is calculated by the maximum lane width occupied by the path; the narrower the lane width, the smaller C3.

[0043] In one implementation, the initial population is iteratively updated according to a first preset number of iterations to obtain the judgment population. For example, the first preset number of iterations is 30. The iterative update is achieved through a velocity update formula and a position update formula, wherein the velocity update formula is: v i m+1 =ω1*v i m +ω2*r1*P best +ω3*r2*G best , where v i m Let ω be the velocity of particle i in the m-th iteration, ω be the inertia weight, ω2 and ω3 be learning factors, and r1 and r2 be random numbers in the range [0,1]. best G represents the optimal position of particle i. best The optimal particle is determined by the following parameters: inertia weight (0.3-0.8) controls the particle's tendency to maintain its original velocity; learning factors (ω2 and ω3) control the weight of the particle's movement towards the optimal individual / group position; random numbers are used to increase the randomness of the search; position update formula: x i m+1 =x i m +v i m+1 , where x i m Let v be the position of particle i in the m-th iteration. i m+1 Let be the velocity of particle i in the (m+1)th iteration; after updating, it is necessary to check whether the parameters are still within the constraint range. If they are outside the range, the parameters are truncated to the boundary to ensure that the parameters are feasible.

[0044] In one implementation, the set of suboptimal particles in each iteration contains multiple suboptimal particles. A suboptimal particle is the best individual of any particle in the iteration process. For example, if there is a particle i (not the best particle), its fitness is retained in each iteration process, and the parameter combination corresponding to the best fitness of that particle is also retained.

[0045] Based on the same inventive concept, this invention also provides an intelligent parking space navigation system for new energy commercial vehicles. See also Figure 2 , Figure 2 A schematic diagram of a parking space intelligent navigation system for new energy commercial vehicles provided in an embodiment of the present invention includes: The data projection module is used to project point cloud data collected by any radar sensor onto a two-dimensional plane to obtain a projected image set. The image acquisition module is used to acquire images from each vision sensor to obtain a visual image set; The image fusion module is used to fuse the projected image and visual image corresponding to each sensor group to obtain a fused feature map; The labeling module is used to input the fused feature map into the object detection model to obtain feature labels, and to add the feature labels to the fused feature map to obtain the target feature map; The condition construction module is used to determine the target feature map set, construct the parking environment based on the target feature map set, and determine the target constraints based on the parking environment; the target feature map set contains target feature maps from any angle collected by new energy commercial vehicles; The path optimization module is used to optimize parking space navigation based on constraints to obtain the target path, and then perform parking space navigation based on the target path.

[0046] The intelligent parking navigation system for new energy commercial vehicles provided by this invention improves the efficiency and accuracy of the detection model by projecting point cloud data onto a two-dimensional plane and fusing it with visual images. The input of the fused feature map provides more precise constraints for parking path optimization, shortens planning time, and increases the success rate of parking paths.

[0047] In one embodiment, the data projection module includes: The coordinate transformation module is used to determine the target spherical coordinates of the target point cloud in the point cloud data, and transform the target spherical coordinates to obtain the target planar coordinates; the target point cloud is any one of the points in the point cloud data; The image stitching module is used to stitch together an initial projected image based on the coordinates of each plane, and to perform coordinate transformation on the initial projected image to obtain a final projected image.

[0048] In one embodiment, the fusion process of the fused feature maps includes: The projected image is subjected to target operations a preset number of times to obtain a projected feature map, and the visual image is input into the patch embedding module to obtain a first feature map. The first feature map is then input into the target attention module to obtain a second feature map. The second feature map is sequentially input into the patch fusion module and the target attention module to obtain the third feature map. The third feature map is sequentially input into the patch fusion module and the target attention module to obtain the fourth feature map. The fourth feature map is sequentially input into the patch fusion module and the target attention module to obtain the fifth feature map. The projected feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map are input into the context aggregation module to obtain the fused feature map.

[0049] In one embodiment, the condition building module includes: The initial condition determination module is used to acquire the physical properties of the new energy commercial vehicle, determine the basic constraints based on the physical properties, and determine the boundary constraints based on the parking environment; the physical properties include: minimum turning radius, maximum curvature, and vehicle planar profile; The final condition determination module is used to determine the target constraint condition based on the basic constraint condition and the boundary constraint condition.

[0050] In one embodiment, the path optimization module includes: The first execution module is used to determine the parameter combination according to the constraint conditions, use the parameter combination as particles, and generate an initial population according to the target constraint conditions; the initial population contains multiple particles. The second execution module is used to evaluate the fitness of the target particles in the initial population and select the particle with the lowest fitness as the optimal particle; the target particle is any one of the particles in the initial population. The third execution module is used to iteratively update the initial population according to a first preset number of iterations to obtain a judgment population, and to determine the set of suboptimal particles in each generation according to the first number of iterations; the set of suboptimal particles in each generation contains multiple suboptimal particles, and the suboptimal particle is the best individual of any particle in the iteration process. The fourth execution module is used to perform update operations on each suboptimal particle in the previous suboptimal particle set to obtain the target suboptimal particle set, determine the target optimal particle in the judgment population, and combine the target optimal particle and the target suboptimal particle set to obtain the target population. The fifth execution module is used to iteratively update the target population until a second preset number of iterations is reached, or if the change in fitness between the current generation's best particle and the previous generation's best particle is less than a preset threshold, then the optimal parameter combination is output, and the target path is determined based on the optimal parameter combination.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A parking space intelligent navigation method for new energy commercial vehicles, characterized in that, New energy commercial vehicles are equipped with multiple sensor groups, each including a vision sensor and a radar sensor, with a one-to-one correspondence between the data acquisition angles of the vision sensor and the radar sensor; the method includes: Point cloud data collected by any radar sensor is projected onto a two-dimensional plane to obtain a projected image set; A visual image set is obtained by acquiring images from each visual sensor; The projected image and visual image corresponding to each sensor group are fused to obtain a fused feature map; The fused feature map is input into the target detection model to obtain feature labels, and the feature labels are added to the fused feature map to obtain the target feature map; A target feature map set is determined, a parking environment is constructed based on the target feature map set, and target constraints are determined based on the parking environment; the target feature map set includes target feature maps from any angle collected by new energy commercial vehicles; The target path is obtained by optimizing parking space navigation based on the constraints, and parking space navigation is performed based on the target path.

2. The intelligent parking space navigation method for new energy commercial vehicles according to claim 1, characterized in that, Projecting the point cloud data onto a two-dimensional plane to obtain a projected image includes: Determine the target spherical coordinates of the target point cloud in the point cloud data, and transform the target spherical coordinates to obtain the target planar coordinates; the target point cloud is any one of the points in the point cloud data; An initial projected image is obtained by stitching together the coordinates of each plane, and the initial projected image is then transformed to obtain the final projected image.

3. The intelligent parking space navigation method for new energy commercial vehicles according to claim 1, characterized in that, The projected image and the visual image are fused to obtain a fused feature map, including: The projected image is subjected to target operations a preset number of times to obtain a projected feature map, and the visual image is input into the patch embedding module to obtain a first feature map. The first feature map is then input into the target attention module to obtain a second feature map. The second feature map is sequentially input into the patch fusion module and the target attention module to obtain the third feature map. The third feature map is sequentially input into the patch fusion module and the target attention module to obtain the fourth feature map. The fourth feature map is sequentially input into the patch fusion module and the target attention module to obtain the fifth feature map. The projected feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map are input into the context aggregation module to obtain the fused feature map.

4. The intelligent parking space navigation method for new energy commercial vehicles according to claim 1, characterized in that, The target constraints are determined based on the parking environment, including: The physical properties of the new energy commercial vehicle are obtained, basic constraints are determined based on the physical properties, and boundary constraints are determined based on the parking environment; the physical properties include: minimum turning radius, maximum curvature, and vehicle planar profile; The target constraints are determined based on the basic constraints and the boundary constraints.

5. The intelligent parking space navigation method for new energy commercial vehicles according to claim 1, characterized in that, The target path is obtained by optimizing parking space navigation based on the aforementioned constraints, including: The parameter combination is determined according to the constraints, and the parameter combination is used as a particle to generate an initial population according to the target constraints; the initial population contains multiple particles. The fitness of the target particles in the initial population is evaluated to obtain the fitness, and the particle with the lowest fitness is taken as the optimal particle; the target particle can be any one of the particles in the initial population. The initial population is iteratively updated according to the first preset number of iterations to obtain the judgment population, and the set of suboptimal particles in each generation is determined according to the first number of iterations; the set of suboptimal particles in each generation contains multiple suboptimal particles, and the suboptimal particle is the best individual of any particle in the iteration process. An update operation is performed on each suboptimal particle in the previous suboptimal particle set to obtain a target suboptimal particle set. The target optimal particle in the judgment population is determined, and the target optimal particle and the target suboptimal particle set are combined to obtain the target population. The target population is iteratively updated until a second preset number of iterations is reached, or if the change in fitness between the current generation's best particle and the previous generation's best particle is less than a preset threshold, then the optimal parameter combination is output, and the target path is determined based on the optimal parameter combination.

6. A parking space intelligent navigation system for new energy commercial vehicles, characterized in that, The system includes: The data projection module is used to project point cloud data collected by any radar sensor onto a two-dimensional plane to obtain a projected image set. The image acquisition module is used to acquire images from each vision sensor to obtain a visual image set; The image fusion module is used to fuse the projected image and visual image corresponding to each sensor group to obtain a fused feature map; The labeling module is used to input the fused feature map into the target detection model to obtain feature labels, and to add the feature labels to the fused feature map to obtain the target feature map; The condition construction module is used to determine the target feature map set, construct a parking environment based on the target feature map set, and determine the target constraint conditions based on the parking environment; the target feature map set includes target feature maps from any angle collected by new energy commercial vehicles; The path optimization module is used to optimize parking space navigation according to the constraints to obtain the target path, and then perform parking space navigation according to the target path.

7. A parking space intelligent navigation system for new energy commercial vehicles according to claim 6, characterized in that, The data projection module includes: The coordinate transformation module is used to determine the target spherical coordinates of the target point cloud in the point cloud data, and transform the target spherical coordinates to obtain the target planar coordinates; the target point cloud is any one of the points in the point cloud data; The image stitching module is used to stitch together an initial projected image based on the coordinates of each plane, and to perform coordinate transformation on the initial projected image to obtain a final projected image.

8. A parking space intelligent navigation system for new energy commercial vehicles according to claim 6, characterized in that, The feature map fusion process includes: The projected image is subjected to target operations a preset number of times to obtain a projected feature map, and the visual image is input into the patch embedding module to obtain a first feature map. The first feature map is then input into the target attention module to obtain a second feature map. The second feature map is sequentially input into the patch fusion module and the target attention module to obtain the third feature map. The third feature map is sequentially input into the patch fusion module and the target attention module to obtain the fourth feature map. The fourth feature map is sequentially input into the patch fusion module and the target attention module to obtain the fifth feature map. The projected feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map are input into the context aggregation module to obtain the fused feature map.

9. A parking space intelligent navigation system for new energy commercial vehicles according to claim 6, characterized in that, The condition construction module includes: The initial condition determination module is used to acquire the physical properties of the new energy commercial vehicle, determine the basic constraints based on the physical properties, and determine the boundary constraints based on the parking environment; the physical properties include: minimum turning radius, maximum curvature, and vehicle planar profile; The final condition determination module is used to determine the target constraint condition based on the basic constraint condition and the boundary constraint condition.

10. A parking space intelligent navigation system for new energy commercial vehicles according to claim 6, characterized in that, The path optimization module includes: The first execution module is used to determine the parameter combination according to the constraint conditions, use the parameter combination as particles, and generate an initial population according to the target constraint conditions; the initial population contains multiple particles. The second execution module is used to evaluate the fitness of the target particles in the initial population and select the particle with the lowest fitness as the optimal particle; the target particle is any one of the particles in the initial population. The third execution module is used to iteratively update the initial population according to a first preset number of iterations to obtain a judgment population, and to determine the set of suboptimal particles in each generation according to the first number of iterations; the set of suboptimal particles in each generation contains multiple suboptimal particles, and the suboptimal particle is the best individual of any particle in the iteration process. The fourth execution module is used to perform update operations on each suboptimal particle in the previous suboptimal particle set to obtain the target suboptimal particle set, determine the target optimal particle in the judgment population, and combine the target optimal particle and the target suboptimal particle set to obtain the target population. The fifth execution module is used to iteratively update the target population until a second preset number of iterations is reached, or if the change in fitness between the current generation's best particle and the previous generation's best particle is less than a preset threshold, then the optimal parameter combination is output, and the target path is determined based on the optimal parameter combination.

Images