A warehouse feature learning method and system based on laser SLAM

Abstract

The present invention discloses a warehouse feature learning method and system based on laser SLAM. It relates to the field of SLAM map feature analysis. Through laser and visual acquisition technology, a cloud point map model is constructed, and warehouse area division and robot positioning are performed. By identifying the landmark features of the defect area, a GAN feature learning model is constructed, and adversarial generation training is performed by fusing laser and visual features to generate simulated map data. If the loss function converges, the contour shape features of the simulated map data are extracted, and the map is updated based on the simulated point cloud data, thereby realizing feature learning of the warehouse environment, improving the navigation accuracy and environmental adaptability of the warehouse robot, and realizing real-time and efficient construction of the warehouse map.

Description

Technical Field

[0001] The present invention relates to the field of SLAM map feature analysis, and more specifically, to a warehouse feature learning method and system based on laser SLAM. Background Art

[0002] In the field of warehousing and logistics, autonomous navigation and precise positioning of warehousing robots is one of the key technologies. Traditional methods rely on GPS or preset paths, but are easily affected by the complexity and dynamic changes of the warehousing environment. In recent years, SLAM (Simultaneous Localization and Mapping) technology has attracted much attention because it can build maps and achieve positioning in real time in unknown environments. However, it is difficult to achieve high-precision environmental understanding and robot navigation by relying solely on laser or visual acquisition methods. At the same time, in the existing technology, the real-time application of SLAM maps is prone to inaccurate positioning data, low map feature quantity, and low efficiency of robot position and posture recognition. In addition, the existing technology lacks methods for real-time map feature analysis and learning, making it difficult to achieve efficient map construction and precise positioning applications, and it is difficult to improve the navigation accuracy and environmental adaptability of warehousing robots, resulting in low efficiency in warehousing management. Summary of the invention

[0003] The present invention overcomes the defects of the prior art and proposes a warehouse feature learning method and system based on laser SLAM.

[0004] The first aspect of the present invention provides a storage feature learning method based on laser SLAM, comprising:

[0005] In the target warehouse robot equipment, the storage environment features of the target area are collected through the laser collection device and the visual collection device to obtain the laser collection features and the visual collection features. The map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM;

[0006] Through the cloud point map model, based on the preset spatial interval, the cloud point map model is divided into multiple storage areas. Combined with the visual acquisition features, the target storage robot is positioned and analyzed in real time. The periodic positioning deviation of the target storage robot is calculated in each storage area, and defective areas with movement deviations are screened out.

[0007] Based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format. Based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set.

[0008] Construct a GAN-based feature learning model, extract the main features of the first map feature set based on the autoencoder to obtain a main feature data set, fuse the main feature data set with the second map feature set and import them into the feature learning model for adversarial generation training, generate simulated map data through the trained feature learning model, and judge the difference between the simulated map data and the main feature data set based on the preset loss function;

[0009] If the preset loss function converges to the preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through the edge detection operator to obtain the contour shape features. Combined with the coordinates of the cloud point map, the contour shape features are mapped to the map point cloud data to generate simulated point cloud data. The simulated point cloud data is imported into the cloud point map model for map update.

[0010] In this solution, in the target warehouse robot equipment, the storage environment features of the target area are collected by a laser collection device and a visual collection device to obtain laser collection features and visual collection features. A map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM, specifically:

[0011] In the target warehouse robot equipment, a laser acquisition device and a visual acquisition device are set;

[0012] The laser collection device includes a laser sensor, a transmitting unit, a scanning unit and a receiving unit;

[0013] The visual acquisition device includes a high-definition camera device and an image storage device;

[0014] The storage environment characteristics of the target area are collected by using a laser collection device and a visual collection device to obtain laser collection characteristics and visual collection characteristics respectively;

[0015] The laser acquisition features are used to analyze the point cloud data, and the point cloud is registered according to the point cloud data to build a cloud point map model based on SLAM.

[0016] In this solution, the cloud point map model is used to divide the area based on the preset spatial interval to form multiple storage areas. The target storage robot is positioned and analyzed in real time in combination with the visual acquisition features. The periodic positioning deviation of the target storage robot is calculated in each storage area, and the defective areas with movement deviation are screened out. Specifically:

[0017] Set the preset space interval, and divide the overall storage map into regions from the cloud point map model to form multiple storage areas, and ensure that the space size of the storage area is within the preset range;

[0018] Through the real-time cloud point data collected by the target warehouse robot, combined with the cloud point map model to perform spatial positioning and path analysis, combined with the visual acquisition features, the target warehouse robot is subjected to auxiliary positioning analysis to generate the real-time positioning data of the target warehouse robot;

[0019] Set N moving cycles, take a storage area as the analysis unit, analyze and determine the deviation between the target storage robot and the target path. The deviation is comprehensively analyzed based on the average speed and the average path deviation distance to generate N deviations.

[0020] Based on the time dimension, N deviations are serialized to form a deviation sequence. Based on the linear regression algorithm, the deviation sequence is evaluated for linear changes to generate a linear change rate. If the linear change rate exceeds the expected change rate range, the corresponding storage area is marked as a defective area.

[0021] Analyze multiple storage areas to filter out all defective areas.

[0022] In this solution, based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format. Based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set, which is specifically:

[0023] Based on the laser acquisition features corresponding to the defect area, point cloud data is extracted in the cloud point map model to form a point cloud data set;

[0024] Through the cloud point map model, landmarks are identified and located in the defect area, and landmark features are extracted in the point cloud data set. The extracted landmark point cloud data is converted into a spatial image, and combined with the coordinate system of the cloud point map model, a first map feature set based on the image format is generated;

[0025] Based on landmark positioning, the corresponding image data is acquired from the visual acquisition features to obtain the initial image data. The contour features and shape features of the landmark objects are extracted from the initial image data based on the Prewitt operator to obtain the second map feature set.

[0026] In this solution, the feature learning model based on GAN is constructed, the main features of the first map feature set are extracted based on the autoencoder to obtain the main feature data set, the main feature data set is feature-fused with the second map feature set and imported into the feature learning model for adversarial generation training, and simulated map data is generated by the trained feature learning model. Based on the preset loss function, the difference between the simulated map data and the main feature data set is judged, specifically:

[0027] Construct a GAN-based feature learning model, which includes a generation model and a discriminant model;

[0028] By using an autoencoder, a low-dimensional space is constructed, the first map feature set is mapped into a low-dimensional data representation, and the low-dimensional data is reconstructed based on a multi-layer perceptron to extract main feature data, and the reconstructed data is used as the main feature data set;

[0029] Based on the feature weighted fusion method, the main feature data set is fused with the second map feature set, and the obtained fused feature set is imported into the feature learning model as the real data;

[0030] The generative model learns features through real data and cyclically generates virtual feature data;

[0031] Import the virtual feature data into the discriminant model for data discrimination. In the generation and discrimination process, optimize the generation model and discriminant model parameters based on the error back propagation algorithm.

[0032] The model adversarial learning is repeated until the generative model and the discriminative model reach Nash equilibrium, completing the model training process.

[0033] The trained feature learning model is used to generate a preset amount of simulated map data, and the mean square error loss function is used to determine the difference between the simulated map data and the main feature data set.

[0034] In this solution, if the preset loss function converges to a preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through an edge detection operator to obtain contour shape features, and the contour shape features are mapped to the map point cloud data in combination with the coordinates of the cloud point map to generate simulated point cloud data, and the simulated point cloud data is imported into the cloud point map model for map update, specifically:

[0035] Determine whether the mean square error loss function converges to a preset range;

[0036] If so, edge detection and contour and shape feature extraction are performed on the simulated map data based on the Prewitt operator to obtain contour shape features;

[0037] Import the contour shape features into the cloud point map model, map and convert the point cloud data through the map coordinates, and generate simulated point cloud data;

[0038] The simulated point cloud data is imported into the cloud point map model as supplementary point cloud data to expand and update the map data.

[0039] In this solution, the cloud point map model performs user interaction through a preset terminal device.

[0040] In this solution, the preset terminal devices include mobile terminal devices and computer terminal devices.

[0041] The second aspect of the present invention further provides a storage feature learning system based on laser SLAM, wherein the storage feature learning program based on laser SLAM is included in the memory, and when the storage feature learning program based on laser SLAM is executed by the processor, the following steps are implemented:

[0042] In the target warehouse robot equipment, the storage environment features of the target area are collected through the laser collection device and the visual collection device to obtain the laser collection features and the visual collection features. The map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM;

[0043] Through the cloud point map model, based on the preset spatial interval, the cloud point map model is divided into multiple storage areas. Combined with the visual acquisition features, the target storage robot is positioned and analyzed in real time. The periodic positioning deviation of the target storage robot is calculated in each storage area, and defective areas with movement deviations are screened out.

[0044] Based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format. Based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set.

[0045] Construct a GAN-based feature learning model, extract the main features of the first map feature set based on the autoencoder to obtain a main feature data set, fuse the main feature data set with the second map feature set and import them into the feature learning model for adversarial generation training, generate simulated map data through the trained feature learning model, and judge the difference between the simulated map data and the main feature data set based on the preset loss function;

[0046] If the preset loss function converges to the preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through the edge detection operator to obtain the contour shape features. Combined with the coordinates of the cloud point map, the contour shape features are mapped to the map point cloud data to generate simulated point cloud data. The simulated point cloud data is imported into the cloud point map model for map update.

[0047] The third aspect of the present invention also provides a computer-readable storage medium, which includes a warehouse feature learning program based on laser SLAM. When the warehouse feature learning program based on laser SLAM is executed by a processor, the steps of the warehouse feature learning method based on laser SLAM as described in any one of the above are implemented.

[0048] The present invention discloses a warehouse feature learning method and system based on laser SLAM. Through laser and visual acquisition technology, a cloud point map model is constructed, and warehouse area division and robot positioning are performed. By identifying the landmark features of the defect area, a GAN feature learning model is constructed, and adversarial generation training is performed by integrating laser and visual features to generate simulated map data. If the loss function converges, the contour shape features of the simulated map data are extracted, and the map is updated based on the simulated point cloud data, thereby realizing feature learning of the warehouse environment, improving the navigation accuracy and environmental adaptability of the warehouse robot, and realizing real-time and efficient construction of the warehouse map. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a storage feature learning method based on laser SLAM of the present invention is shown;

[0050] Figure 2 A block diagram of a warehouse feature learning system based on laser SLAM of the present invention is shown. DETAILED DESCRIPTION

[0051] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0053] Figure 1 A flow chart of a warehouse feature learning method based on laser SLAM of the present invention is shown.

[0054] like Figure 1 As shown, the first aspect of the present invention provides a storage feature learning method based on SLAM, comprising:

[0055] S102, in the target warehouse robot equipment, the storage environment features of the target area are collected by a laser collection device and a visual collection device to obtain laser collection features and visual collection features, and a map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM;

[0056] S104, using the cloud point map model, based on the preset spatial interval, the cloud point map model is divided into regions to form multiple storage areas, and the target storage robot is positioned and moved in real time in combination with the visual acquisition features, and the periodic positioning deviation of the target storage robot is calculated in each storage area, and defective areas with movement deviations are screened out;

[0057] S106, based on the laser acquisition features corresponding to the defect area, perform landmark recognition and landmark feature extraction in the cloud point map model to generate a first map feature set based on the image format, and based on the landmark positioning, perform landmark feature extraction from the visual acquisition features to obtain a second map feature set;

[0058] S108, constructing a feature learning model based on GAN, extracting main features of the first map feature set based on the autoencoder to obtain a main feature data set, fusing the main feature data set with the second map feature set and importing the data into the feature learning model for adversarial generation training, generating simulated map data through the trained feature learning model, and judging the difference between the simulated map data and the main feature data set based on a preset loss function;

[0059] S110, if the preset loss function converges to a preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through an edge detection operator to obtain contour shape features, and the contour shape features are mapped to the map point cloud data in combination with the coordinates of the cloud point map to generate simulated point cloud data, and the simulated point cloud data is imported into the cloud point map model for map updating.

[0060] It is worth mentioning here that in the existing technology, the real-time application of SLAM maps is prone to inaccurate positioning data, low map feature quantity, and low efficiency in robot position and posture recognition, especially for early data collection and analysis, map construction and navigation tasks. The existing technology lacks real-time map feature analysis and learning methods, making it difficult to achieve efficient map construction and precise positioning applications, and difficult to improve the navigation accuracy and environmental adaptability of warehouse robots, resulting in inefficient warehouse management.

[0061] In the present invention, a two-dimensional spatial map feature collection is performed by a laser collection device and a visual collection device, and a SLAM cloud point map model is first constructed. The movement deviation of the warehouse robot is periodically analyzed to mark the defective area. For the defective area, the corresponding point cloud feature data and the visual feature data are fused and analyzed. With the landmark feature as the core, the fused feature is subjected to feature learning and simulation generation. The generated simulation data is judged and imported into the cloud point map model for real-time feature supplementation and updating, thereby realizing the environmental feature learning and dynamic map updating of the SALM map model, effectively improving the recognition accuracy and navigation accuracy of the landmarks in the warehouse map, further improving the environmental adaptability of the warehouse robot, and realizing the efficient construction of the cloud point map.

[0062] According to an embodiment of the present invention, in the target warehouse robot device, the storage environment features of the target area are collected by a laser collection device and a visual collection device to obtain laser collection features and visual collection features, and a map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM, specifically:

[0063] In the target warehouse robot equipment, a laser acquisition device and a visual acquisition device are set;

[0064] The laser collection device includes a laser sensor, a transmitting unit, a scanning unit and a receiving unit;

[0065] The visual acquisition device includes a high-definition camera device and an image storage device;

[0066] The storage environment characteristics of the target area are collected by using a laser collection device and a visual collection device to obtain laser collection characteristics and visual collection characteristics respectively;

[0067] The laser acquisition features are used to analyze the point cloud data, and the point cloud is registered according to the point cloud data to build a cloud point map model based on SLAM.

[0068] It should be noted that the target warehouse robot equipment includes assisting warehouse robots, AGVs, forklift robots, etc.

[0069] According to an embodiment of the present invention, the cloud point map model is used to divide the area based on the preset spatial interval to form multiple storage areas, and the target storage robot is positioned and moved in real time in combination with the visual acquisition features. The periodic positioning deviation of the target storage robot is calculated in each storage area, and the defective areas with movement deviation are screened out, specifically:

[0070] Set the preset space interval, and divide the overall storage map into regions from the cloud point map model to form multiple storage areas, and ensure that the space size of the storage area is within the preset range;

[0071] Through the real-time cloud point data collected by the target warehouse robot, combined with the cloud point map model to perform spatial positioning and path analysis, combined with the visual acquisition features, the target warehouse robot is subjected to auxiliary positioning analysis to generate the real-time positioning data of the target warehouse robot;

[0072] Set N moving cycles, take a storage area as the analysis unit, analyze and determine the deviation between the target storage robot and the target path. The deviation is comprehensively analyzed based on the average speed and the average path deviation distance to generate N deviations.

[0073] Based on the time dimension, N deviations are serialized to form a deviation sequence. Based on the linear regression algorithm, the deviation sequence is evaluated for linear changes to generate a linear change rate. If the linear change rate exceeds the expected change rate range, the corresponding storage area is marked as a defective area.

[0074] Analyze multiple storage areas to filter out all defective areas.

[0075] It should be noted that the preset spatial interval is specifically a spatial distance value. The real-time cloud point data is obtained according to the laser acquisition device. The linear change rate can effectively reflect the path deviation trend of the target storage robot.

[0076] According to an embodiment of the present invention, based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format, and based on the landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set, which is specifically:

[0077] Based on the laser acquisition features corresponding to the defect area, point cloud data is extracted in the cloud point map model to form a point cloud data set;

[0078] Through the cloud point map model, landmarks are identified and located in the defect area, and landmark features are extracted in the point cloud data set. The extracted landmark point cloud data is converted into a spatial image, and combined with the coordinate system of the cloud point map model, a first map feature set based on the image format is generated;

[0079] Based on landmark positioning, the corresponding image data is acquired from the visual acquisition features to obtain the initial image data. The contour features and shape features of the landmark objects are extracted from the initial image data based on the Prewitt operator to obtain the second map feature set.

[0080] It should be noted that the landmarks include identification points or areas that can be used for positioning and target storage robot movement analysis. The first map feature set is a feature set converted into image format by performing spatial object analysis through landmark point cloud data, which reflects landmark features in image form.

[0081] According to an embodiment of the present invention, the feature learning model based on GAN is constructed, the main features of the first map feature set are extracted based on the autoencoder to obtain a main feature data set, the main feature data set is feature-fused with the second map feature set and imported into the feature learning model for adversarial generation training, simulated map data is generated by the trained feature learning model, and the difference between the simulated map data and the main feature data set is judged based on a preset loss function, specifically:

[0082] Construct a GAN-based feature learning model, which includes a generation model and a discriminant model;

[0083] By using an autoencoder, a low-dimensional space is constructed, the first map feature set is mapped into a low-dimensional data representation, and the low-dimensional data is reconstructed based on a multi-layer perceptron to extract main feature data, and the reconstructed data is used as the main feature data set;

[0084] Based on the feature weighted fusion method, the main feature data set is fused with the second map feature set, and the obtained fused feature set is imported into the feature learning model as the real data;

[0085] The generative model learns features through real data and cyclically generates virtual feature data;

[0086] Import the virtual feature data into the discriminant model for data discrimination. In the generation and discrimination process, optimize the generation model and discriminant model parameters based on the error back propagation algorithm.

[0087] The model adversarial learning is repeated until the generative model and the discriminative model reach Nash equilibrium, completing the model training process.

[0088] The trained feature learning model is used to generate a preset amount of simulated map data, and the mean square error loss function is used to determine the difference between the simulated map data and the main feature data set.

[0089] It should be noted that the feature learning model is used to learn landmark features collected based on laser and landmark features collected by machine vision technology, and the features are fused for learning and simulation feature generation.

[0090] Landmark features are important environmental feature data in the map model, and they play an important role in the positioning analysis, deviation analysis, and posture analysis of warehouse robots.

[0091] According to an embodiment of the present invention, if the preset loss function converges to a preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through an edge detection operator to obtain contour shape features, and the contour shape features are mapped to map point cloud data in combination with the coordinates of the cloud point map to generate simulated point cloud data, and the simulated point cloud data is imported into the cloud point map model for map update, specifically:

[0092] Determine whether the mean square error loss function converges to a preset range;

[0093] If so, edge detection and contour and shape feature extraction are performed on the simulated map data based on the Prewitt operator to obtain contour shape features;

[0094] Import the contour shape features into the cloud point map model, map and convert the point cloud data through the map coordinates, and generate simulated point cloud data;

[0095] The simulated point cloud data is imported into the cloud point map model as supplementary point cloud data to expand and update the map data.

[0096] According to an embodiment of the present invention, the cloud point map model performs user interaction through a preset terminal device.

[0097] It should be noted that the user interaction includes the user checking the map and the marking analysis of the point cloud map.

[0098] According to an embodiment of the present invention, the preset terminal device includes a mobile terminal device and a computer terminal device.

[0099] According to an embodiment of the present invention, the method of fusing the main feature data set with the second map feature set and importing the result into a feature learning model for adversarial generation training further includes:

[0100] For a defect area, the main feature data set is extracted based on multiple target warehouse robots, and the corresponding collected visual angle is recorded for each target warehouse machine to form multiple main feature data sets;

[0101] Each main feature dataset corresponds to a visual angle;

[0102] Marking the main feature data sets corresponding to two different visual angles as a first feature data set and a second feature data set;

[0103] Based on the DCA fusion algorithm, the first feature data set and the second feature data set are respectively used to generate the first feature map and the second feature map, and the class activation map is obtained by multiplying the feature map with the corresponding class weight, and the class activation map is weighted fused to obtain a set of fused feature sets;

[0104] Based on the combination of every two visual angles, multiple groups of fused feature sets are generated, and the multiple groups of fused feature sets are sequentially imported into the feature learning model for adversarial generation training to obtain multiple groups of simulated map data.

[0105] It should be noted that the category weights are preset weights.

[0106] It is worth mentioning here that in the case of multi-target warehouse robots, the same landmark or the same area can often collect feature data from multiple angles. Compared with the single-angle collection features of a target warehouse robot, the feature data collected from multiple angles has better feature learning value, and the generated simulation data is better applied and fits the real map features. Therefore, in the case of multi-target warehouse robots, the present invention is based on multi-angle features and combines multiple feature groups as real data in a DCA feature fusion manner. When generating simulation data, simulation generation is performed based on multiple feature group data. Multiple groups of simulation data are generated corresponding to the simulation data, and the preset loss function is used to evaluate whether the multiple groups of simulation data meet expectations. The embodiment of the present invention can effectively improve the accuracy and availability of real-time map modeling during the operation of the warehouse robot by rationally utilizing the data collected by multiple robots for fusion analysis.

[0107] According to an embodiment of the present invention, it also includes:

[0108] One defect area is used as the unit of analysis;

[0109] After the cloud point map model is updated, K movement cycles are set;

[0110] In K moving cycles, the deviation between the target storage robot and the target path is analyzed and determined to obtain K deviations;

[0111] Based on the time dimension, K deviations are serialized to obtain the current sequence, and the current sequence is linearly fitted and evaluated based on the linear regression equation to obtain the fitting equation and the current linear change rate;

[0112] The deviation of multiple moving periods is predicted by fitting equations to generate predicted deviation;

[0113] The navigation accuracy of the cloud point map model and the map update effect feedback are evaluated based on the current linear change rate and the predicted deviation. If the current linear change rate is within the expected change rate range and the predicted deviation is lower than the preset deviation, the corresponding defective area will be marked as a general storage area.

[0114] It should be noted that the present invention performs positioning accuracy feedback evaluation for map updating by setting K moving cycles, and determines whether to mark the defective area as a normal area according to the linear fitting deviation.

[0115] N and K are preset values ​​and can be changed based on actual analysis.

[0116] Figure 2 A block diagram of a warehouse feature learning system based on laser SLAM of the present invention is shown.

[0117] The second aspect of the present invention further provides a storage feature learning system 2 based on laser SLAM, the system comprising: a memory 21, a processor 22, the memory 21 comprising a storage feature learning program based on laser SLAM, the storage feature learning program based on laser SLAM when executed by the processor 22 implements the following steps:

[0118] In the target warehouse robot equipment, the storage environment features of the target area are collected through the laser collection device and the visual collection device to obtain the laser collection features and the visual collection features. The map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM;

[0119] Through the cloud point map model, based on the preset spatial interval, the cloud point map model is divided into multiple storage areas. Combined with the visual acquisition features, the target storage robot is positioned and analyzed in real time. The periodic positioning deviation of the target storage robot is calculated in each storage area, and defective areas with movement deviations are screened out.

[0120] Based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format. Based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set.

[0121] Construct a GAN-based feature learning model, extract the main features of the first map feature set based on the autoencoder to obtain a main feature data set, fuse the main feature data set with the second map feature set and import them into the feature learning model for adversarial generation training, generate simulated map data through the trained feature learning model, and judge the difference between the simulated map data and the main feature data set based on the preset loss function;

[0122] If the preset loss function converges to the preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through the edge detection operator to obtain the contour shape features. Combined with the coordinates of the cloud point map, the contour shape features are mapped to the map point cloud data to generate simulated point cloud data. The simulated point cloud data is imported into the cloud point map model for map update.

[0123] It is worth mentioning here that in the existing technology, the real-time application of SLAM maps is prone to inaccurate positioning data, low map feature quantity, and low efficiency in robot position and posture recognition, especially for early data collection and analysis, map construction and navigation tasks. The existing technology lacks real-time map feature analysis and learning methods, making it difficult to achieve efficient map construction and precise positioning applications, and difficult to improve the navigation accuracy and environmental adaptability of warehouse robots, resulting in inefficient warehouse management.

[0124] In the present invention, a two-dimensional spatial map feature collection is performed by a laser collection device and a visual collection device, and a SLAM cloud point map model is first constructed. The movement deviation of the warehouse robot is periodically analyzed to mark the defective area. For the defective area, the corresponding point cloud feature data and the visual feature data are fused and analyzed. With the landmark feature as the core, the fused feature is subjected to feature learning and simulation generation. The generated simulation data is judged and imported into the cloud point map model for real-time feature supplementation and updating, thereby realizing the environmental feature learning and dynamic map updating of the SALM map model, effectively improving the recognition accuracy and navigation accuracy of the landmarks in the warehouse map, further improving the environmental adaptability of the warehouse robot, and realizing the efficient construction of the cloud point map.

[0125] According to an embodiment of the present invention, in the target warehouse robot device, the storage environment features of the target area are collected by a laser collection device and a visual collection device to obtain laser collection features and visual collection features, and a map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM, specifically:

[0126] In the target warehouse robot equipment, a laser acquisition device and a visual acquisition device are set;

[0127] The laser collection device includes a laser sensor, a transmitting unit, a scanning unit and a receiving unit;

[0128] The visual acquisition device includes a high-definition camera device and an image storage device;

[0129] The storage environment characteristics of the target area are collected by using a laser collection device and a visual collection device to obtain laser collection characteristics and visual collection characteristics respectively;

[0130] The laser acquisition features are used to analyze the point cloud data, and the point cloud is registered according to the point cloud data to build a cloud point map model based on SLAM.

[0131] It should be noted that the target warehouse robot equipment includes assisting warehouse robots, AGVs, forklift robots, etc.

[0132] According to an embodiment of the present invention, the cloud point map model is used to divide the area based on the preset spatial interval to form multiple storage areas, and the target storage robot is positioned and moved in real time in combination with the visual acquisition features. The periodic positioning deviation of the target storage robot is calculated in each storage area, and the defective areas with movement deviation are screened out, specifically:

[0133] Set the preset space interval, and divide the overall storage map into regions from the cloud point map model to form multiple storage areas, and ensure that the space size of the storage area is within the preset range;

[0134] Through the real-time cloud point data collected by the target warehouse robot, combined with the cloud point map model to perform spatial positioning and path analysis, combined with the visual acquisition features, the target warehouse robot is subjected to auxiliary positioning analysis to generate the real-time positioning data of the target warehouse robot;

[0135] Set N moving cycles, take a storage area as the analysis unit, analyze and determine the deviation between the target storage robot and the target path. The deviation is comprehensively analyzed based on the average speed and the average path deviation distance to generate N deviations.

[0136] Based on the time dimension, N deviations are serialized to form a deviation sequence. Based on the linear regression algorithm, the deviation sequence is evaluated for linear changes to generate a linear change rate. If the linear change rate exceeds the expected change rate range, the corresponding storage area is marked as a defective area.

[0137] Analyze multiple storage areas to filter out all defective areas.

[0138] It should be noted that the preset spatial interval is specifically a spatial distance value. The real-time cloud point data is obtained according to the laser acquisition device. The linear change rate can effectively reflect the path deviation trend of the target storage robot.

[0139] According to an embodiment of the present invention, based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format, and based on the landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set, which is specifically:

[0140] Based on the laser acquisition features corresponding to the defect area, point cloud data is extracted in the cloud point map model to form a point cloud data set;

[0141] Through the cloud point map model, landmarks are identified and located in the defect area, and landmark features are extracted in the point cloud data set. The extracted landmark point cloud data is converted into a spatial image, and combined with the coordinate system of the cloud point map model, a first map feature set based on the image format is generated;

[0142] Based on landmark positioning, the corresponding image data is acquired from the visual acquisition features to obtain the initial image data. The contour features and shape features of the landmark objects are extracted from the initial image data based on the Prewitt operator to obtain the second map feature set.

[0143] It should be noted that the landmarks include identification points or areas that can be used for positioning and target storage robot movement analysis. The first map feature set is a feature set converted into image format by performing spatial object analysis through landmark point cloud data, which reflects landmark features in image form.

[0144] According to an embodiment of the present invention, the feature learning model based on GAN is constructed, the main features of the first map feature set are extracted based on the autoencoder to obtain a main feature data set, the main feature data set is feature-fused with the second map feature set and imported into the feature learning model for adversarial generation training, simulated map data is generated by the trained feature learning model, and the difference between the simulated map data and the main feature data set is judged based on a preset loss function, specifically:

[0145] Construct a GAN-based feature learning model, which includes a generation model and a discriminant model;

[0146] By using an autoencoder, a low-dimensional space is constructed, the first map feature set is mapped into a low-dimensional data representation, and the low-dimensional data is reconstructed based on a multi-layer perceptron to extract main feature data, and the reconstructed data is used as the main feature data set;

[0147] Based on the feature weighted fusion method, the main feature data set is fused with the second map feature set, and the obtained fused feature set is imported into the feature learning model as the real data;

[0148] The generative model learns features through real data and cyclically generates virtual feature data;

[0149] Import the virtual feature data into the discriminant model for data discrimination. In the generation and discrimination process, optimize the generation model and discriminant model parameters based on the error back propagation algorithm.

[0150] The model adversarial learning is repeated until the generative model and the discriminative model reach Nash equilibrium, completing the model training process.

[0151] The trained feature learning model is used to generate a preset amount of simulated map data, and the mean square error loss function is used to determine the difference between the simulated map data and the main feature data set.

[0152] It should be noted that the feature learning model is used to learn landmark features collected based on laser and landmark features collected by machine vision technology, and the features are fused for learning and simulation feature generation.

[0153] Landmark features are important environmental feature data in the map model, and they play an important role in the positioning analysis, deviation analysis, and posture analysis of warehouse robots.

[0154] According to an embodiment of the present invention, if the preset loss function converges to a preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through an edge detection operator to obtain contour shape features, and the contour shape features are mapped to map point cloud data in combination with the coordinates of the cloud point map to generate simulated point cloud data, and the simulated point cloud data is imported into the cloud point map model for map update, specifically:

[0155] Determine whether the mean square error loss function converges to a preset range;

[0156] If so, edge detection and contour and shape feature extraction are performed on the simulated map data based on the Prewitt operator to obtain contour shape features;

[0157] Import the contour shape features into the cloud point map model, map and convert the point cloud data through the map coordinates, and generate simulated point cloud data;

[0158] The simulated point cloud data is imported into the cloud point map model as supplementary point cloud data to expand and update the map data.

[0159] According to an embodiment of the present invention, the cloud point map model performs user interaction through a preset terminal device.

[0160] It should be noted that the user interaction includes the user checking the map and the marking analysis of the point cloud map.

[0161] According to an embodiment of the present invention, the preset terminal device includes a mobile terminal device and a computer terminal device.

[0162] The third aspect of the present invention also provides a computer-readable storage medium, which includes a warehouse feature learning program based on laser SLAM. When the warehouse feature learning program based on laser SLAM is executed by a processor, the steps of the warehouse feature learning method based on laser SLAM as described in any one of the above are implemented.

[0163] The present invention discloses a warehouse feature learning method and system based on laser SLAM. Through laser and visual acquisition technology, a cloud point map model is constructed, and warehouse area division and robot positioning are performed. By identifying the landmark features of the defect area, a GAN feature learning model is constructed, and adversarial generation training is performed by integrating laser and visual features to generate simulated map data. If the loss function converges, the contour shape features of the simulated map data are extracted, and the map is updated based on the simulated point cloud data, thereby realizing feature learning of the warehouse environment, improving the navigation accuracy and environmental adaptability of the warehouse robot, and realizing real-time and efficient construction of the warehouse map.

[0164] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0165] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0166] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0167] A person of ordinary skill in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0168] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0169] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A warehouse feature learning method based on laser SLAM, characterized in that: include: In the target warehouse robot equipment, the storage environment features of the target area are collected through the laser collection device and the visual collection device to obtain the laser collection features and the visual collection features. The map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM; Through the cloud point map model, based on the preset spatial interval, the cloud point map model is divided into multiple storage areas. Combined with the visual acquisition features, the target storage robot is positioned and analyzed in real time. The periodic positioning deviation of the target storage robot is calculated in each storage area, and defective areas with movement deviations are screened out. Based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format. Based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set. Construct a GAN-based feature learning model, extract the main features of the first map feature set based on the autoencoder to obtain a main feature data set, fuse the main feature data set with the second map feature set and import them into the feature learning model for adversarial generation training, generate simulated map data through the trained feature learning model, and judge the difference between the simulated map data and the main feature data set based on the preset loss function; If the preset loss function converges to the preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through the edge detection operator to obtain the contour shape features. Combined with the coordinates of the cloud point map, the contour shape features are mapped to the map point cloud data to generate simulated point cloud data. The simulated point cloud data is imported into the cloud point map model for map update.

2. According to the warehouse feature learning method based on laser SLAM according to claim 1, it is characterized in that: In the target warehouse robot equipment, the storage environment features of the target area are collected by the laser collection device and the visual collection device to obtain the laser collection features and the visual collection features. The map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM, specifically: In the target warehouse robot equipment, a laser acquisition device and a visual acquisition device are set; The laser collection device includes a laser sensor, a transmitting unit, a scanning unit and a receiving unit; The visual acquisition device includes a high-definition camera device and an image storage device; The storage environment characteristics of the target area are collected by using a laser collection device and a visual collection device to obtain laser collection characteristics and visual collection characteristics respectively; The laser acquisition features are used to analyze the point cloud data, and the point cloud is registered according to the point cloud data to build a cloud point map model based on SLAM.

3. According to a warehouse feature learning method based on laser SLAM according to claim 1, it is characterized in that: The cloud point map model is used to divide the area based on the preset spatial interval to form multiple storage areas. The target storage robot is positioned and analyzed in real time in combination with the visual acquisition features. The periodic positioning deviation of the target storage robot is calculated in each storage area, and the defective areas with movement deviation are screened out. Specifically: Set the preset space interval, and divide the overall storage map into regions from the cloud point map model to form multiple storage areas, and ensure that the space size of the storage area is within the preset range; Through the real-time cloud point data collected by the target warehouse robot, combined with the cloud point map model to perform spatial positioning and path analysis, combined with the visual acquisition features, the target warehouse robot is subjected to auxiliary positioning analysis to generate the real-time positioning data of the target warehouse robot; Set N moving cycles, take a storage area as the analysis unit, analyze and determine the deviation between the target storage robot and the target path. The deviation is comprehensively analyzed based on the average speed and the average path deviation distance to generate N deviations. Based on the time dimension, N deviations are serialized to form a deviation sequence. Based on the linear regression algorithm, the deviation sequence is evaluated for linear changes to generate a linear change rate. If the linear change rate exceeds the expected change rate range, the corresponding storage area is marked as a defective area. Analyze multiple storage areas to filter out all defective areas.

4. The storage feature learning method based on laser SLAM according to claim 1 is characterized in that: Based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format, and based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set, specifically: Based on the laser acquisition features corresponding to the defect area, point cloud data is extracted in the cloud point map model to form a point cloud data set; Through the cloud point map model, landmarks are identified and located in the defect area, and landmark features are extracted in the point cloud data set. The extracted landmark point cloud data is converted into a spatial image, and combined with the coordinate system of the cloud point map model, a first map feature set based on the image format is generated; Based on landmark positioning, the corresponding image data is acquired from the visual acquisition features to obtain the initial image data. The contour features and shape features of the landmark objects are extracted from the initial image data based on the Prewitt operator to obtain the second map feature set.

5. The method for learning storage features based on laser SLAM according to claim 1, characterized in that: The feature learning model based on GAN is constructed, the main features of the first map feature set are extracted based on the autoencoder to obtain the main feature data set, the main feature data set is feature-fused with the second map feature set and imported into the feature learning model for adversarial generation training, simulated map data is generated by the trained feature learning model, and the difference between the simulated map data and the main feature data set is judged based on the preset loss function, specifically: Construct a GAN-based feature learning model, which includes a generation model and a discriminant model; By using an autoencoder, a low-dimensional space is constructed, the first map feature set is mapped into a low-dimensional data representation, and the low-dimensional data is reconstructed based on a multi-layer perceptron to extract main feature data, and the reconstructed data is used as the main feature data set; Based on the feature weighted fusion method, the main feature data set is fused with the second map feature set, and the obtained fused feature set is imported into the feature learning model as the real data; The generative model learns features through real data and cyclically generates virtual feature data; Import the virtual feature data into the discriminant model for data discrimination. In the generation and discrimination process, optimize the generation model and discriminant model parameters based on the error back propagation algorithm. The model adversarial learning is repeated until the generative model and the discriminative model reach Nash equilibrium, completing the model training process. The trained feature learning model is used to generate a preset amount of simulated map data, and the mean square error loss function is used to determine the difference between the simulated map data and the main feature data set.

6. A warehouse feature learning method based on laser SLAM according to claim 5, characterized in that: If the preset loss function converges to the preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through an edge detection operator to obtain contour shape features, and the contour shape features are mapped to the map point cloud data in combination with the coordinates of the cloud point map to generate simulated point cloud data, and the simulated point cloud data is imported into the cloud point map model for map update, specifically: Determine whether the mean square error loss function converges to a preset range; If so, edge detection and contour and shape feature extraction are performed on the simulated map data based on the Prewitt operator to obtain contour shape features; Import the contour shape features into the cloud point map model, map and convert the point cloud data through the map coordinates, and generate simulated point cloud data; The simulated point cloud data is imported into the cloud point map model as supplementary point cloud data to expand and update the map data.

7. The method for learning storage features based on laser SLAM according to claim 1, characterized in that: The cloud point map model performs user interaction through a preset terminal device.

8. The method for learning storage features based on laser SLAM according to claim 7, characterized in that: The preset terminal equipment includes a mobile terminal equipment and a computer terminal equipment.

9. A warehouse feature learning system based on laser SLAM, characterized in that: The system includes: a memory and a processor, wherein the memory includes a storage feature learning program based on laser SLAM, and when the storage feature learning program based on laser SLAM is executed by the processor, the following steps are implemented: In the target warehouse robot equipment, the storage environment features of the target area are collected through the laser collection device and the visual collection device to obtain the laser collection features and the visual collection features. The map is constructed in real time based on the laser collection features to form a cloud point map model based on SLAM; Through the cloud point map model, based on the preset spatial interval, the cloud point map model is divided into multiple storage areas. Combined with the visual acquisition features, the target storage robot is positioned and analyzed in real time. The periodic positioning deviation of the target storage robot is calculated in each storage area, and defective areas with movement deviations are screened out. Based on the laser acquisition features corresponding to the defect area, landmark recognition and landmark feature extraction are performed in the cloud point map model to generate a first map feature set based on the image format. Based on landmark positioning, landmark features are extracted from the visual acquisition features to obtain a second map feature set. Construct a GAN-based feature learning model, extract the main features of the first map feature set based on the autoencoder to obtain a main feature data set, fuse the main feature data set with the second map feature set and import them into the feature learning model for adversarial generation training, generate simulated map data through the trained feature learning model, and judge the difference between the simulated map data and the main feature data set based on the preset loss function; If the preset loss function converges to the preset range, the simulated map data is subjected to edge detection and contour and shape feature extraction through the edge detection operator to obtain the contour shape features. Combined with the coordinates of the cloud point map, the contour shape features are mapped to the map point cloud data to generate simulated point cloud data. The simulated point cloud data is imported into the cloud point map model for map update.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a laser SLAM-based warehouse feature learning program. When the laser SLAM-based warehouse feature learning program is executed by a processor, the steps of the laser SLAM-based warehouse feature learning method as described in any one of claims 1 to 8 are implemented.

Images