Scene coding and coding matching method and device for underground parking lot, and electronic equipment

Abstract

The invention provides a scene coding and code matching method and device for an underground parking lot, and electronic equipment, and the method comprises the steps: carrying out the rough matching through a random walk descriptor, determining a candidate mapping topological interval matched with a positioning topological interval, carrying out the fine matching through a shape context descriptor, and obtaining a candidate mapping topological interval matched with a positioning topological interval; and finally, obtaining a target matching result of matching the first target shape context descriptor of the positioning topological interval with the second target shape context descriptor of the candidate mapping topological interval, so that the matching workload of subsequent fine matching is greatly reduced in the rough matching process, and the matching efficiency of overall matching is improved. And the accuracy of the fine matching process is good, the matching between the points is centimeter-level, and the cross validation effect is also formed by fusing the two matching modes, so that the scene recognition result is more accurate and robust.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular, to a method, device, and electronic device for scene encoding and encoding matching in an underground parking lot. Background Art

[0002] The valet parking technology for underground scenes is one of the important applications of autonomous driving technology. The scene recognition technology in underground parking lots becomes particularly crucial.

[0003] Existing scene recognition (i.e., matching) methods mainly rely on traditional image processing and machine learning technologies. The descriptors extracted by traditional technologies have poor robustness in complex environments, especially in situations with large light changes, severe occlusions, or lack of texture information, and the performance is not satisfactory.

[0004] In recent years, some deep learning-based methods have begun to be applied to scene recognition tasks. Although these methods have improved the recognition accuracy in some aspects, there are still problems such as high consumption of computing resources and poor real-time performance. In addition, these methods have high requirements for the amount of data and are prone to overfitting in the case of insufficient data.

[0005] In summary, in the process of scene recognition, how to balance the efficiency (i.e., real-time performance) and accuracy of scene matching has become a technical problem that needs to be solved urgently at present. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method, device, and electronic device for scene encoding and encoding matching in an underground parking lot to alleviate the technical problem that in the existing scene recognition process, the efficiency (i.e., real-time performance) and accuracy of scene matching cannot be balanced.

[0007] In a first aspect, an embodiment of the present invention provides a method for scene encoding and encoding matching in an underground parking lot, including:

[0008] Obtain multiple frames of panoramic semantic segmentation maps within a topological interval, and construct a topological node map corresponding to the topological interval according to the target clustering center points extracted from the multiple frames of panoramic semantic segmentation maps, where the topological interval is a preset distance range, and the topological interval includes: a plurality of mapping topological intervals and a positioning topological interval;

[0009] Extract the random walk descriptor of the topological node map according to the topological node map;

[0010] Extract a preset number of target panoramic semantic segmentation maps at equal intervals from the multiple frames of panoramic semantic segmentation maps within the topological interval, and extract the corresponding shape context descriptors according to the target panoramic semantic segmentation maps;

[0011] Match the random walk descriptors corresponding to the positioning topological interval with the random walk descriptors corresponding to each of the mapping topological intervals respectively, and determine a candidate mapping topological interval that matches the positioning topological interval in the mapping topological interval according to the first matching result;

[0012] Match each shape context descriptor corresponding to the positioning topological interval with each shape context descriptor corresponding to each of the candidate mapping topological intervals, and determine a target matching result in which the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate mapping topological interval according to the second matching result.

[0013] Further, obtain multiple frames of omnidirectional semantic segmentation maps within the topological interval, including:

[0014] Within the topological interval, obtain parking lot images collected by an on-vehicle omnidirectional camera, and perform projection stitching on each obtained parking lot image to obtain multiple frames of bird's-eye view images within the topological interval;

[0015] Perform semantic segmentation on each frame of the bird's-eye view image to obtain multiple frames of omnidirectional semantic segmentation maps within the topological interval, where the categories of the semantic segmentation include: parking lines, solid lines, dashed lines, columns, speed bumps;

[0016] Construct a topological node graph corresponding to the topological interval according to the target clustering center points extracted from multiple frames of the omnidirectional semantic segmentation maps, including:

[0017] Extract the target clustering center points from multiple frames of the omnidirectional semantic segmentation maps, where the target clustering center points include: column clustering center points, parking space corner point clustering center points, arrow clustering center points, speed bump clustering center points;

[0018] Convert the target clustering center points to the world coordinate system to obtain topological nodes, and further obtain a topological node graph composed of the topological nodes.

[0019] Further, extract the random walk descriptors of the topological node graph according to the topological node graph, including:

[0020] Determine non-independent points and independent points among the topological nodes in the topological node graph according to a first preset distance threshold, where the non-independent points are the topological nodes in the topological node graph that have other topological nodes within the first preset distance threshold from themselves;

[0021] Determine the dimension of the random walk descriptors of the non-independent points, where the dimension is K*1 dimension, K is the nth power of m, m represents the number of categories of all non-independent points in the topological node graph, and n represents the step length of the random walk;

[0022] Traverse the non-independent points, and determine each neighbor point of the current non-independent point among the topological nodes in the topological node graph according to the step size of the random walk and the first preset distance threshold. Among them, the neighbor points include: direct neighbor points and / or indirect neighbor points. The direct neighbor points are topological nodes in the topological node graph whose distance from the current non-independent point is less than the first preset distance threshold. When the indirect neighbor points are multi-level indirect neighbor points, each level of indirect neighbor points is a topological node in the topological node graph whose distance from the previous level of indirect neighbor points is less than the first preset distance threshold. When the indirect neighbor points are first-level indirect neighbor points, the previous level of indirect neighbor points is the direct neighbor point;

[0023] Calculate the index of the category combination according to the index calculation formula: the category of the current non-independent point * m to the power of n - 1 + the category of the direct neighbor point * m to the power of n - 2 + the category of the first-level indirect neighbor point * m to the power of n - 3 + the category of the second-level indirect neighbor point * m to the power of n - 4 + … + the category of the last-level indirect neighbor point * m to the power of n - n. Among them, the category combination is: the category of the current non-independent point, the category of the direct neighbor point, the category of the first-level indirect neighbor point, the category of the second-level indirect neighbor point, …, the category of the last-level indirect neighbor point;

[0024] Increment the element at the position of the random walk descriptor corresponding to the index by 1, thereby obtaining the random walk descriptor of the current non-independent point, and obtain the non-independent point random walk descriptor according to the random walk descriptors of all non-independent points;

[0025] Obtain the random walk descriptor of the independent points, and determine the random walk descriptor of the topological node graph according to the non-independent point random walk descriptor and the random walk descriptor of the independent points. Among them, the dimension of the random walk descriptor of the independent points is the same as the dimension of the random walk descriptor of the non-independent points.

[0026] Further, extract the corresponding shape context descriptor according to the target panoramic semantic segmentation map, including:

[0027] Extract the contour points of the target line with the category of line in the target panoramic semantic segmentation map;

[0028] Downsample the contour points to obtain the downsampled contour points;

[0029] For each of the downsampled contour points, construct a polar logarithmic coordinate system centered on it. Among them, the polar logarithmic coordinate system includes: P angles and Q distances, thereby obtaining P * Q grid regions;

[0030] Mapping other downsampled contour points around the central downsampled contour point to the P*Q grid areas, and counting the number of downsampled contour points in each of the grid areas;

[0031] Normalizing the number of downsampled contour points in each of the grid areas to obtain a normalized number of downsampled contour points in each of the grid areas;

[0032] A shape context descriptor of the downsampled contour points of the center is generated according to the normalized quantity, and a shape context descriptor corresponding to the target surround semantic segmentation map is determined according to the shape context descriptors of all the downsampled contour points of the center.

[0033] Further, the random walk descriptor corresponding to the positioning topology interval is matched with the random walk descriptor corresponding to each of the mapping topology intervals, and a candidate mapping topology interval matching the positioning topology interval is determined in the mapping topology interval according to the first matching result, including:

[0034] Constructing a scoring matrix, wherein the number of rows of the scoring matrix is ​​the number of topological nodes in the positioning topological interval, the number of columns of the scoring matrix is ​​the number of topological nodes in the current mapping topological interval, and the value of each element in the scoring matrix is ​​0;

[0035] Traversing the random walk descriptor of each row of the positioning topology interval;

[0036] If the topological node corresponding to the random walk descriptor of the current row of the positioning topological interval is an independent point, skipping the random walk descriptor of the current row of the positioning topological interval;

[0037] If the topological node corresponding to the random walk descriptor of the current row of the positioning topological interval is a non-independent point, traverse the random walk descriptors of each row of the current mapping topological interval;

[0038] If the category of the topological node corresponding to the random walk descriptor of the current row of the positioning topology interval is different from the category of the topological node corresponding to the random walk descriptor of the current row of the current mapping topology interval, skip the random walk descriptor of the current row of the current mapping topology interval;

[0039] If the topological node corresponding to the random walk descriptor of the current row of the positioning topology interval is of the same category as the topological node corresponding to the random walk descriptor of the current row of the current mapping topology interval, then the matching similarity between the random walk descriptor of the current row of the positioning topology interval and the random walk descriptor of the current row of the current mapping topology interval is calculated;

[0040] Determine whether the random walk descriptor of the current row of the positioning topology interval is proportional to the random walk descriptor of the current row of the current mapping topology interval;

[0041] If they are proportional, set the matching similarity to 0, and fill the matching similarity of 0 into the position in the scoring matrix corresponding to the current row of the positioning topology interval and the current row of the current mapping topology interval;

[0042] If they are not proportional, fill the matching similarity into the position in the scoring matrix corresponding to the current row of the positioning topology interval and the current row of the current mapping topology interval until the traversal of the positioning topology interval and the current mapping topology interval is completed, thereby obtaining a target scoring matrix, where each element in the target scoring matrix represents the similarity between the topological node of the positioning topology interval corresponding to its corresponding row and the topological node of the current mapping topology interval corresponding to its corresponding column;

[0043] Determine the first optimal matching point pair and the corresponding target rotation transformation and target translation transformation according to the target scoring matrix;

[0044] Traverse the topological nodes of the positioning topology interval, and transform the current topological node of the positioning topology interval by using the target rotation transformation and the target translation transformation to obtain the transformed current topological node;

[0045] Traverse the topological nodes of the current mapping topology interval;

[0046] If the category of the current topological node of the positioning topology interval is different from that of the current topological node of the current mapping topology interval, skip the current topological node of the current mapping topology interval;

[0047] If the distance between the transformed current topological node and the current topological node of the current mapping topology interval is greater than the second preset distance threshold, skip the current topological node of the current mapping topology interval;

[0048] If the current topological node of the positioning topology interval is a non-independent point, or the current topological node of the current mapping topology interval is a non-independent point, skip the current topological node of the current mapping topology interval;

[0049] Take the current topological node of the positioning topology interval and the current topological node of the current mapping topology interval that are not skipped as the second optimal matching point pair;

[0050] Calculate the matching success ratio based on the first optimal matching point pair and the second optimal matching point pair;

[0051] If the matching success ratio is greater than the first preset ratio threshold, then use the current mapping topology interval as the candidate mapping topology interval of the positioning topology interval.

[0052] Further, determining the first optimal matching point pair and the corresponding target rotation transformation and target translation transformation according to the target score matrix includes:

[0053] Determine target elements in the target score matrix whose element values are greater than the preset similarity threshold, and determine candidate matching point pairs according to the target elements;

[0054] Randomly select a target number of candidate matching point pairs from the candidate matching point pairs, use the topological nodes belonging to the positioning topology interval among the target number of candidate matching point pairs as target points, and use the topological nodes belonging to the current mapping topology interval among the target number of candidate matching point pairs as source points;

[0055] Calculate the rotation transformation and translation transformation between the source points and the target points according to the target number of candidate matching point pairs;

[0056] Use the rotation transformation and the translation transformation to transform each target point to obtain the transformed target points;

[0057] Determine whether the transformed target points are included within the third preset distance threshold of each source point;

[0058] If included, increment the number of matching points by 1 to obtain the target number of matching points;

[0059] If the target number of matching points is less than the preset quantity threshold, determine that the target number of candidate matching point pairs are incorrect matching point pairs, and return to execute the process of randomly selecting a target number of candidate matching point pairs from the candidate matching point pairs until the target number of matching points is greater than the preset quantity threshold;

[0060] If the target number of matching points is greater than the preset quantity threshold, determine that the target number of candidate matching point pairs are the first optimal matching point pair, and use the corresponding rotation transformation and translation transformation as the target rotation transformation and the target translation transformation.

[0061] Further, match each shape context descriptor corresponding to the positioning topology interval with each shape context descriptor corresponding to each candidate mapping topology interval, and determine the target matching result of the first target shape context descriptor corresponding to the positioning topology interval and the second target shape context descriptor corresponding to the candidate mapping topology interval according to the second matching result, including:

[0062] Calculate the similarity matrix between each shape context descriptor corresponding to the positioning topology interval and each shape context descriptor corresponding to each of the candidate mapping topology intervals using the chi-square formula;

[0063] Process each of the similarity matrices using the Hungarian matching strategy to obtain multiple matching point pair results;

[0064] Determine that the first target shape context descriptor of the positioning topology interval and the second target shape context descriptor of the candidate mapping topology interval match according to the target matching point pair result with the largest number of matches among the multiple matching point pair results, where the first target shape context descriptor and the second target shape context descriptor are the shape context descriptors corresponding to the target matching point pair result.

[0065] In a second aspect, an embodiment of the present invention further provides a scene coding and coding matching device for an underground parking lot, including:

[0066] An acquisition and construction unit, configured to acquire multiple frames of panoramic semantic segmentation maps within a topology interval, and construct a topology node map corresponding to the topology interval according to the target clustering center points extracted from the multiple frames of panoramic semantic segmentation maps, where the topology interval is a preset distance range, and the topology interval includes: multiple mapping topology intervals and one positioning topology interval;

[0067] A random walk descriptor extraction unit, configured to extract a random walk descriptor of the topology node map according to the topology node map;

[0068] A shape context descriptor extraction unit, configured to equally spaced extract a preset number of target panoramic semantic segmentation maps from multiple frames of panoramic semantic segmentation maps within the topology interval, and extract corresponding shape context descriptors according to the target panoramic semantic segmentation maps;

[0069] A first matching unit, configured to match the random walk descriptor corresponding to the positioning topology interval with the random walk descriptor corresponding to each of the mapping topology intervals, and determine a candidate mapping topology interval that matches the positioning topology interval among the mapping topology intervals according to the first matching result;

[0070] A second matching unit, configured to match each shape context descriptor corresponding to the positioning topology interval with each shape context descriptor corresponding to each of the candidate mapping topology intervals, and determine a target matching result that the first target shape context descriptor of the positioning topology interval matches the second target shape context descriptor of the candidate mapping topology interval according to the second matching result.

[0071] In a third aspect, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method according to any one of the above first aspects are implemented.

[0072] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium storing machine-executable instructions. When the machine-executable instructions are called and run by a processor, the machine-executable instructions cause the processor to run the method according to any one of the above first aspects.

[0073] In an embodiment of the present invention, a method for scene encoding and encoding matching of an underground parking lot is provided, including: obtaining multiple frames of omnidirectional vision semantic segmentation maps within a topological interval, and constructing a topological node map corresponding to the topological interval according to the target clustering center points extracted from the multiple frames of omnidirectional vision semantic segmentation maps, where the topological interval is a preset distance range, and the topological interval includes: multiple map-building topological intervals and a positioning topological interval; extracting random walk descriptors of the topological node map according to the topological node map; extracting a preset number of target omnidirectional vision semantic segmentation maps at equal intervals from the multiple frames of omnidirectional vision semantic segmentation maps within the topological interval, and extracting corresponding shape context descriptors according to the target omnidirectional vision semantic segmentation maps; respectively matching the random walk descriptors corresponding to the positioning topological interval with the random walk descriptors corresponding to each map-building topological interval, and determining a candidate map-building topological interval that matches the positioning topological interval in the map-building topological intervals according to the first matching result; matching each shape context descriptor corresponding to the positioning topological interval with each shape context descriptor corresponding to each candidate map-building topological interval, and determining a target matching result in which the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate map-building topological interval according to the second matching result. As can be seen from the above description, in the method for scene encoding and encoding matching of the underground parking lot of the present invention, first, a candidate map-building topological interval that matches the positioning topological interval is determined in the map-building topological intervals through random walk descriptors, and then a target matching result in which the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate map-building topological interval is determined through shape context descriptors, that is, first, a rough match is performed through random walk descriptors to determine a candidate map-building topological interval that matches the positioning topological interval, and then a fine match is performed through shape context descriptors to finally obtain a target matching result in which the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate map-building topological interval. The above process of rough matching greatly reduces the matching workload of subsequent fine matching, improves the matching efficiency of the overall matching, and the above process of fine matching has good accuracy, and the matching between points is centimeter-level. By fusing the above two matching methods, a cross-validation effect is also formed, making the result of scene recognition more accurate and robust, and alleviating the technical problem that in the traditional scene recognition process, it is impossible to balance the efficiency (i.e., real-time performance) and accuracy of scene matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0075] Figure 1 Flow chart of a method for scene encoding and encoding matching of an underground parking lot provided by an embodiment of the present invention;

[0076] Figure 2 Schematic diagram of a topological node graph provided by an embodiment of the present invention;

[0077] Figure 3 Schematic diagram of extracting a topological node combination provided by an embodiment of the present invention;

[0078] Figure 4 Schematic diagram of a target line provided by an embodiment of the present invention;

[0079] Figure 5 Schematic diagram of downsampled contour points provided by an embodiment of the present invention;

[0080] Figure 6 Schematic diagram of a pole-pair coordinate system provided by an embodiment of the present invention;

[0081] Figure 7 Schematic diagram of optimal matching point pairs provided by an embodiment of the present invention;

[0082] Figure 8 Schematic diagram of a result of a matching point pair provided by an embodiment of the present invention;

[0083] Figure 9 Schematic diagram of a device for scene encoding and encoding matching of an underground parking lot provided by an embodiment of the present invention;

[0084] Figure 10 Schematic diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0085] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the embodiments. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0086] In the traditional scene recognition process, it is impossible to take into account both the efficiency (i.e., real-time performance) and accuracy of scene matching.

[0087] Based on this, in the scenario encoding and encoding matching method for the underground parking lot of the present invention, first, a candidate mapping topological interval that matches the positioning topological interval is determined in the mapping topological interval through the random walk descriptor. Then, the target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate mapping topological interval is determined through the shape context descriptor. That is, first, a rough match is performed through the random walk descriptor to determine a candidate mapping topological interval that matches the positioning topological interval. Then, a fine match is performed through the shape context descriptor. Finally, the target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate mapping topological interval is obtained. The above process of rough matching greatly reduces the matching workload of subsequent fine matching, improves the matching efficiency of the overall matching, and the above process of fine matching has good accuracy, and the matching between points is at the centimeter level. By fusing the above two matching methods, a cross-verification effect is also formed, making the result of scene recognition more accurate and robust.

[0088] To facilitate the understanding of this embodiment, first, a detailed introduction to a scenario encoding and encoding matching method for an underground parking lot disclosed in an embodiment of the present invention is provided.

[0089] Embodiment 1:

[0090] According to an embodiment of the present invention, an embodiment of a scenario encoding and encoding matching method for an underground parking lot is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0091] Figure 1 is a flowchart of a scenario encoding and encoding matching method for an underground parking lot according to an embodiment of the present invention, as Figure 1 shown, the method includes the following steps:

[0092] Step S102, obtain multiple frames of panoramic semantic segmentation maps within the topological interval, and construct a topological node map corresponding to the topological interval according to the target clustering center points extracted from the multiple frames of panoramic semantic segmentation maps, where the topological interval is a preset distance range, and the topological interval includes: multiple mapping topological intervals and one positioning topological interval;

[0093] Specifically, the above topological interval is a preset distance range. For example, during mapping, every 30 meters is a mapping topological interval. A topological interval includes multiple frames of panoramic semantic segmentation maps, and one topological interval corresponds to one topological node map.

[0094] Step S104, extracting a random walk descriptor of the topological node graph according to the topological node graph;

[0095] The random walk descriptor of the topological node graph can represent the information of the local structure between the topological nodes in the topological node graph.

[0096] Step S106, extracting a preset number of target surround view semantic segmentation maps at equal intervals from the multi-frame surround view semantic segmentation maps within the topological interval, and extracting corresponding shape context descriptors according to the target surround view semantic segmentation maps;

[0097] Specifically, as mentioned above, in a mapping topology interval of every 30 meters, a frame of target surround semantic segmentation map can be obtained every 6 meters, and each frame of target surround semantic segmentation map has a corresponding shape context descriptor.

[0098] Step S108, matching the random walk descriptor corresponding to the positioning topology interval with the random walk descriptor corresponding to each mapping topology interval, and determining a candidate mapping topology interval matching the positioning topology interval in the mapping topology interval according to the first matching result;

[0099] Step S110, matching each shape context descriptor corresponding to the positioning topology interval with each shape context descriptor corresponding to each candidate mapping topology interval, and determining a target matching result of matching the first target shape context descriptor of the positioning topology interval with the second target shape context descriptor of the candidate mapping topology interval according to the second matching result.

[0100] The above process is briefly described below in the form of an example:

[0101] Mapping process:

[0102] A random walk descriptor is constructed every 30 meters, and a shape context descriptor is extracted every 6 meters. Finally, multiple mapping topology intervals X={X1, X2, X3, X4, …} are obtained. Each mapping topology interval contains 5 shape context descriptors {X1: Y={y1, y2, y3, y4, y5}}, {X2: Y={y6, y7, y8, y9, y10}}, which are not listed here one by one.

[0103] Positioning process:

[0104] Obtain the random walk descriptor of the current positioning topology interval (CurX), and extract the shape context descriptor every 6 meters during this period. Then, the positioning topology interval contains 5 shape context descriptors {CurX: Cur Y = {Cur y1, Cur y2, Cur y3, Cur y4, Cur y5}};

[0105] Match the random walk descriptors corresponding to the positioning topological intervals with the random walk descriptors corresponding to each mapping topological interval respectively, find the preset number of the most similar random walk descriptors, and correspondingly obtain the preset number of candidate mapping topological intervals;

[0106] Match each shape context descriptor corresponding to the positioning topological interval with each shape context descriptor corresponding to each candidate mapping topological interval, find the shape context descriptor with the largest number of similar matches, which is the target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate mapping topological interval. Subsequently, the precise pose transformation can be solved according to the target matching result of the first target shape context descriptor and the second target shape context descriptor.

[0107] In an embodiment of the present invention, a method for scene encoding and encoding matching of an underground parking lot is provided, including: obtaining multiple frames of panoramic semantic segmentation maps within a topological interval, and constructing a topological node map corresponding to the topological interval according to the target clustering center points extracted from the multiple frames of panoramic semantic segmentation maps, where the topological interval is a preset distance range, and the topological interval includes: multiple map-building topological intervals and a positioning topological interval; extracting random walk descriptors of the topological node map according to the topological node map; extracting a preset number of target panoramic semantic segmentation maps at equal intervals from the multiple frames of panoramic semantic segmentation maps within the topological interval, and extracting corresponding shape context descriptors according to the target panoramic semantic segmentation maps; matching the random walk descriptors corresponding to the positioning topological interval with the random walk descriptors corresponding to each map-building topological interval respectively, and determining a candidate map-building topological interval matching the positioning topological interval in the map-building topological intervals according to the first matching result; matching each shape context descriptor corresponding to the positioning topological interval with each shape context descriptor corresponding to each candidate map-building topological interval respectively, and determining a target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate map-building topological interval according to the second matching result. As can be seen from the above description, in the method for scene encoding and encoding matching of the underground parking lot of the present invention, first, a candidate map-building topological interval matching the positioning topological interval is determined in the map-building topological intervals through random walk descriptors, and then a target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate map-building topological interval is determined through shape context descriptors, that is, first, rough matching is performed through random walk descriptors to determine a candidate map-building topological interval matching the positioning topological interval, and then fine matching is performed through shape context descriptors to finally obtain a target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate map-building topological interval. The above process of rough matching greatly reduces the matching workload of subsequent fine matching, improves the matching efficiency of the overall matching, and the above process of fine matching has good accuracy, and the matching between points is centimeter-level. By fusing the above two matching methods, a cross-verification effect is also formed, making the result of scene recognition more accurate and robust, and alleviating the technical problem that in the traditional scene recognition process, it is impossible to take into account both the efficiency (i.e., real-time performance) and accuracy of scene matching.

[0108] The above content briefly introduces the method for scene encoding and encoding matching of the underground parking lot of the present invention, and the following describes the specific content involved in detail.

[0109] In an optional embodiment of the present invention, obtaining multiple frames of panoramic semantic segmentation maps within a topological interval specifically includes the following steps:

[0110] (1) In the topological interval, obtain the parking lot images collected by the vehicle surround-view camera, and perform projection stitching on the parking lot images obtained each time to obtain multiple bird's-eye view images in the topological interval;

[0111] (2) Perform semantic segmentation on each frame of the bird's-eye view image to obtain multiple surround-view semantic segmentation images in the topological interval, where the categories of semantic segmentation include: parking lines, solid lines, dashed lines, columns, and speed bumps.

[0112] In an alternative embodiment of the present invention, a topological node graph corresponding to the topological interval is constructed according to the target clustering center points extracted from the multiple surround-view semantic segmentation images, which specifically includes the following steps:

[0113] (1) Extract the target clustering center points from the multiple surround-view semantic segmentation images, where the target clustering center points include: column clustering center points, parking space corner point clustering center points, arrow clustering center points, and speed bump clustering center points;

[0114] Specifically, the target points (such as column points) can be first extracted from each frame of the surround-view semantic segmentation image, and then the average value of multiple target points of the same target extracted from the multiple surround-view semantic segmentation images is calculated to obtain the target clustering center point (such as the column clustering center point).

[0115] (2) Convert the target clustering center points to the world coordinate system to obtain topological nodes, and then obtain a topological node graph composed of the topological nodes.

[0116] Specifically, using the self-vehicle pose information provided by the odometer, these extracted target clustering center points are converted to the world coordinate system to obtain topological nodes, and then a topological node graph is generated. As Figure 2 shown, the green ones are the parking space corner points and the red ones are the columns.

[0117] In an alternative embodiment of the present invention, the random walk descriptors of the topological node graph are extracted according to the topological node graph, which specifically includes the following steps:

[0118] (1) Determine the non-independent points and independent points among the topological nodes in the topological node graph according to the first preset distance threshold, where the non-independent points are the topological nodes in which there are other topological nodes within the first preset distance threshold from themselves;

[0119] (2) Determine the dimension of the random walk descriptor of the non-independent points, where the dimension is K*1 dimension, K is the nth power of m, m represents the number of categories of all non-independent points in the topological node graph, and n represents the step length of the random walk;

[0120] (3) Traverse the non - independent points, and determine each neighbor point of the current non - independent point among the topological nodes in the topological node graph according to the step size of the random walk and the first preset distance threshold. Here, the neighbor points include: direct neighbor points and / or indirect neighbor points. The direct neighbor point is a topological node in the topological node graph whose distance from the current non - independent point is less than the first preset distance threshold. When the indirect neighbor point is a multi - level indirect neighbor point, each level of indirect neighbor point is a topological node in the topological node graph whose distance from its upper - level indirect neighbor point is less than the first preset distance threshold. When the indirect neighbor point is a first - level indirect neighbor point, its upper - level indirect neighbor point is the direct neighbor point;

[0121] (4) Calculate the index of the category combination according to the index calculation formula: the category of the current non - independent point * m to the power of n - 1+ the category of the direct neighbor point * m to the power of n - 2+ the category of the first - level indirect neighbor point * m to the power of n - 3+ the category of the second - level indirect neighbor point * m to the power of n - 4+…+ the category of the last - level indirect neighbor point * m to the power of n - n, where the category combination is: the category of the current non - independent point, the category of the direct neighbor point, the category of the first - level indirect neighbor point, the category of the second - level indirect neighbor point, …, the category of the last - level indirect neighbor point;

[0122] (5) Increment the element at the position of the random - walk descriptor corresponding to the index by 1, thereby obtaining the random - walk descriptor of the current non - independent point, and obtain the non - independent point random - walk descriptor according to the random - walk descriptors of all non - independent points;

[0123] (6) Obtain the random - walk descriptor of the independent point, and determine the random - walk descriptor of the topological node graph according to the non - independent point random - walk descriptor and the random - walk descriptor of the independent point. Here, the dimension of the random - walk descriptor of the independent point is the same as that of the random - walk descriptor of the non - independent point.

[0124] The above process is introduced below:

[0125] After obtaining the topological node graph, extract the random - walk descriptor for each topological node in it, and then determine the random - walk descriptor of the topological node graph according to the random - walk descriptors of all topological nodes. The specific steps are as follows:

[0126] Set the first preset distance threshold M, traverse all topological nodes in the topological node graph. If there are other topological nodes within the range of M from itself (i.e., the current topological node), then consider the current topological node as a non - independent point; otherwise, the current topological node is an independent point;

[0127] Construct a K * 1 dimensional matrix Des, which is the random walk descriptor of each non - independent point, and the values of each element in this matrix are initially 0. K is the nth power of m, where m represents the number of categories of all non - independent points in the topological node graph, and n is the step length of the random walk. For example, if there are 11 categories of all non - independent points, given certain numbers (such as the pillar category is 1, the parking space corner point category is 2, etc.), generally from 0 to 10, and the step length of the random walk is 3 topological nodes (including the root node), then K = 1331;

[0128] As Figure 3 shown, traverse the non - independent points, find all the direct neighbor points Ner of the current non - independent point Cur, and save them in P1;

[0129] Traverse all the direct neighbor points found in the previous step, find all the direct neighbor points of this direct neighbor point, which are the first - level indirect neighbor points NerNer of the current non - independent point Cur, and record their categories;

[0130] Further traverse the first - level indirect neighbor points NerNer that meet the conditions (meet the conditions of the step length of the random walk), find all the direct neighbor points of this first - level indirect neighbor point NerNer, which are the second - level indirect neighbor points NerNerNer of the current non - independent point Cur, and record their categories,

[0131] Continue to traverse and search according to the above process until the step length of the random walk is reached;

[0132] Calculate the value of the current non - independent point Cur's category * m to the power of n - 1+ the direct neighbor point's category * m to the power of n - 2+ the first - level indirect neighbor point's category * m to the power of n - 3+ the second - level indirect neighbor point's category * m to the power of n - 4+…+ the category of the last - level indirect neighbor point * m to the power of n - n (denoted as v), and use it as the index of the matrix Des.

[0133] Through the index v, increment the value corresponding to Des[v] by one, indicating that this category combination (i.e., the category combination of [the category of the current non - independent point, the category of the direct neighbor point, the category of the first - level indirect neighbor point, the category of the second - level indirect neighbor point,…, the category of the last - level indirect neighbor point]) has occurred 1 time;

[0134] For example, if there are 11 categories of all non - independent points, the step length of the random walk is 3 nodes. Currently, the category of the current non - independent point Cur is 1, the category of the direct neighbor point Ner is 3, and the category of the first - level indirect neighbor point NerNer is 4. Then v = 1 * 121+3 * 11+4 * 1 = 158. Then increment the value corresponding to Des

[158] by 1, indicating that the category combination (1, 3, 4) has occurred 1 time.

[0135] If the step size of the random walk is 5, then the category combination for each set of calculations is 5 values;

[0136] For an independent point, its random walk descriptor is a 0 matrix of 1331 * 1 dimension (taking 11 categories and a random walk step size of 3 as an example).

[0137] Then, by accumulating the random walk descriptors Des calculated for each topological node to generate a Q * K matrix, the entire random walk descriptor TotalDes of the topological node graph corresponding to this topological interval can be obtained. Q is the number of topological nodes within this topological interval (i.e., the topological node graph).

[0138] In an optional embodiment of the present invention, extracting the corresponding shape context descriptor according to the target panoramic semantic segmentation map specifically includes the following steps:

[0139] (1) Extract the contour points of the target line with the category of line in the target panoramic semantic segmentation map;

[0140] Specifically, the above target line can be a stop line or a solid line, and no specific limitation is imposed here. After extracting the target line, binaryzation processing can be performed. The effective pixel points are the pixel points of the line category, and the rest are invalid pixel points. For the extraction of the target line (i.e., the contour), the canny operator can be used. The obtained target line is as Figure 4 shown, and further the contour points of the target line can be obtained (the target line can be directly used as the contour points).

[0141] (2) Downsample the contour points to obtain the downsampled contour points;

[0142] Specifically, in order to ensure as evenly spaced contour points as possible, the idea of Jitendra sampling is used. The specific steps are as follows:

[0143] Set the total number S of the downsampled contour points;

[0144] Randomly permute all the obtained contour points, so that point pairs can be randomly deleted;

[0145] Calculate the distance between one contour point and each other contour point in the randomly sorted contour point set;

[0146] Find the contour point pair with the smallest distance, and remove one of the contour points (in the contour point pair with the smallest distance). Take the remaining contour points as all the contour points, and return to execute the step of randomly permuting all the obtained contour points until the number of the finally remaining contour points is equal to the set total number S of the downsampled contour points. The schematic diagram of the downsampled contour points is as Figure 5 shown (this figure is not for Figure 4)(obtained after downsampling the contour points of the target line in it).

[0147] (3) For each downsampled contour point, construct a polar logarithm coordinate system centered on it, where the polar logarithm coordinate system includes: P angles and Q distances, and then obtain P*Q grid regions;

[0148] Specifically, as Figure 6 shown, the left figure shows a polar logarithm coordinate system constructed centered on a downsampled contour point, and the right figure shows a polar logarithm coordinate system constructed centered on another downsampled contour point. The red lines in the figure represent the downsampled contour points.

[0149] (4) Map the other downsampled contour points around the central downsampled contour point into the P*Q grid regions, and count the number of downsampled contour points in each grid region;

[0150] (5) Normalize the number of downsampled contour points in each grid region to obtain the normalized number of downsampled contour points in each grid region;

[0151] (6) Generate a shape context descriptor for the central downsampled contour point according to the normalized number, and determine the shape context descriptor corresponding to the target panoramic semantic segmentation map according to the shape context descriptors of all the central downsampled contour points.

[0152] Specifically, the histogram matrix (i.e., the shape context descriptor) extracted from each downsampled contour point is R*1-dimensional, and the size of R is equal to P*Q. After obtaining the normalized number of downsampled contour points in each grid region, the shape context descriptor of the central downsampled contour point is obtained. The R-dimensional histogram matrix (shape context descriptor) of this central downsampled contour point represents the positional relationship between this downsampled contour point and other surrounding downsampled contour points. As Figure 6 shown, the polar logarithm coordinate system has 8 angles and 5 distances, then the dimension R of the shape context descriptor extracted for each downsampled contour point is 8*5 = 40*1 dimension.

[0153] In an optional embodiment of the present invention, match the random walk descriptors corresponding to the positioning topological intervals with the random walk descriptors corresponding to each mapping topological interval respectively, and determine the candidate mapping topological intervals matching the positioning topological intervals in the mapping topological intervals according to the first matching result, which specifically includes the following steps:

[0154] (1) Construct a scoring matrix, where the number of rows of the scoring matrix is the number of topological nodes in the positioning topological interval, the number of columns of the scoring matrix is the number of topological nodes in the current mapping topological interval, and the value of each element in the scoring matrix is 0;

[0155] Specifically, construct a scoring matrix ScoreMatrix. Its number of rows is the number of topological nodes Q1 in the positioning topological interval, and the number of columns is the number of topological nodes Q2 in the current mapping topological interval, and the initial values are all 0. This scoring matrix ScoreMatrix represents the similarity between pairwise topological nodes in two topological intervals.

[0156] (2) Traverse the random walk descriptors of each row in the positioning topological interval;

[0157] (3) If the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval is an independent point, skip the random walk descriptor of the current row in the positioning topological interval;

[0158] (4) If the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval is a non - independent point, traverse the random walk descriptors of each row in the current mapping topological interval;

[0159] (5) If the categories of the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval and the topological node corresponding to the random walk descriptor of the current row in the current mapping topological interval are different, skip the random walk descriptor of the current row in the current mapping topological interval;

[0160] (6) If the categories of the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval and the topological node corresponding to the random walk descriptor of the current row in the current mapping topological interval are the same, calculate the matching similarity between the random walk descriptor of the current row in the positioning topological interval and the random walk descriptor of the current row in the current mapping topological interval;

[0161] Specifically, the processes of (2) - (6) above are introduced again:

[0162] Traverse rows 0 to Q1 in the random walk descriptor TotalDes1 of the positioning topological interval. Each row represents the random walk descriptor of a topological node;

[0163] a. If the topological node corresponding to this row is an independent point, skip the current topological node in the positioning topological interval being traversed;

[0164] b. Traverse rows 0 to Q2 in the random walk descriptor TotalDes2 of the current mapping topological interval;

[0165] ① If the categories of the topological nodes in Q1 and the topological nodes in Q2 are different, skip the current topological node in the current mapping topological interval traversed.

[0166] ② Calculate the matching similarity of the random walk descriptors Des1 and Des2 of the two topological nodes, and cosine similarity can be used for calculation.

[0167] (7) Determine whether the random walk descriptor of the current row in the positioning topological interval is proportional to the random walk descriptor of the current row in the current mapping topological interval;

[0168] (8) If they are proportional, set the matching similarity to 0, and fill the matching similarity of 0 into the position in the scoring matrix corresponding to the current row of the positioning topological interval and the current row of the current mapping topological interval;

[0169] (9) If they are not proportional, fill the matching similarity into the position in the scoring matrix corresponding to the current row of the positioning topological interval and the current row of the current mapping topological interval until the traversal of the positioning topological interval and the current mapping topological interval is completed, thereby obtaining the target scoring matrix, where each element in the target scoring matrix represents the similarity between the topological node in the positioning topological interval corresponding to its corresponding row and the topological node in the current mapping topological interval corresponding to its corresponding column;

[0170] The process of the above (7)-(9) is introduced again below:

[0171] Since the calculation method of cosine similarity gives the same result for proportional vectors, such as (1, 1, 1), (2, 2, 2) and (3, 3, 3), and the cosine similarity results of pairwise calculations are all 1. But for random walk descriptors, their results are required to be different. So it is necessary to additionally judge whether the two random walk descriptors Des1 and Des2 are proportional. If they are proportional, then the matching similarity is 0. If they are not proportional, then directly use cosine similarity to calculate the matching similarity, and finally fill the calculated matching similarity into the ScoreMatrix correspondingly.

[0172] (10) Determine the first optimal matching point pair and the corresponding target rotation transformation and target translation transformation according to the target scoring matrix;

[0173] Specifically, it includes the following steps:

[0174] 1) Determine the target elements in the target scoring matrix whose element values are greater than the preset similarity threshold, and determine the candidate matching point pairs according to the target elements;

[0175] Specifically, the above preset similarity threshold can be 0.8. When determining the candidate matching point pairs, it is determined according to the row and column where the target element is located.

[0176] 2) Randomly select a target number of candidate matching point pairs from the candidate matching point pairs, and use the topological nodes belonging to the positioning topological interval among the target number of candidate matching point pairs as target points, and use the topological nodes belonging to the current mapping topological interval among the target number of candidate matching point pairs as source points;

[0177] 3) Calculate the rotation transformation and translation transformation between the source point and the target point according to the target number of candidate matching point pairs;

[0178] 4) Use the rotation transformation and translation transformation to transform each target point to obtain the transformed target points;

[0179] 5) Determine whether the transformed target points are included within the third preset distance threshold of each source point;

[0180] 6) If included, increment the number of matching points by 1 to obtain the target number of matching points;

[0181] 7) If the target number of matching points is less than the preset quantity threshold, determine that the target number of candidate matching point pairs are incorrect matching point pairs, and return to execute the process of randomly selecting a target number of candidate matching point pairs from the candidate matching point pairs until the target number of matching points is greater than the preset quantity threshold, or the number of loop iterations reaches the preset number of times;

[0182] 8) If the target number of matching points is greater than the preset quantity threshold, determine that the target number of candidate matching point pairs are the first optimal matching point pairs, and use the corresponding rotation transformation and translation transformation as the target rotation transformation and target translation transformation.

[0183] The above process is introduced again as follows:

[0184] (1) Set a preset similarity threshold E, such as 0.8, traverse the target scoring matrix, and only retain the topological node pairs in the target scoring matrix whose similarity scores exceed the preset similarity threshold E as candidate matching point pairs;

[0185] (2) Set a target number H, such as 4, and a preset number of times X, such as 1000, randomly select H candidate matching point pairs from the candidate matching point pairs for inlier checking:

[0186] a. Among the 4 candidate matching point pairs, mark the topological nodes in the positioning topological interval as target points, and mark the topological nodes in the current mapping topological interval as source points;

[0187] b. Use the randomly selected 4 candidate matching point pairs to calculate the rotation transformation C and translation transformation D between the source point and the target point;

[0188] c. Apply the solved rotation transformation and translation transformation to all the target points, aiming to transform the target points to the same coordinate system as the source points;

[0189] d. Set a third preset distance threshold T, such as 0.5, and calculate whether there is a transformed target point within the third preset distance threshold T for each source point. If there is, increment the statistical value Sum of the number of matching points of the inlier by 1 to obtain the final number of target matching points;

[0190] e. If the number of target matching points is lower than a certain proportion (i.e., the preset quantity threshold) of the number of source points, such as 0.8, then consider these 4 candidate matching point pairs as incorrect matching point pairs, exit, and enter the next loop

[0191] f. If the number of target matching points reaches a certain proportion (i.e., the preset quantity threshold) of the number of source points, such as 0.8, then consider these 4 candidate matching point pairs as correct, that is, the first optimal matching point pairs, with relatively high credibility. Record the indices of these first optimal matching point pairs and the solved target rotation transformation and target translation transformation, and consider that the optimal matching has been found, then exit the loop.

[0192] (11) Traverse the topological nodes in the positioning topological interval, and transform the current topological node in the positioning topological interval using the target rotation transformation and target translation transformation to obtain the transformed current topological node;

[0193] (12) Traverse the topological nodes in the current mapping topological interval;

[0194] (13) If the category of the current topological node in the positioning topological interval is different from that of the current topological node in the current mapping topological interval, then skip the current topological node in the current mapping topological interval;

[0195] (14) If the distance between the transformed current topological node and the current topological node in the current mapping topological interval is greater than the second preset distance threshold, then skip the current topological node in the current mapping topological interval;

[0196] (15) If the current topological node in the positioning topological interval is a non - independent point, or the current topological node in the current mapping topological interval is a non - independent point, then skip the current topological node in the current mapping topological interval;

[0197] (16) Take the current topological node in the positioning topological interval and the current topological node in the current mapping topological interval that are not skipped as the second optimal matching point pairs;

[0198] The process of the above (11)-(16) is introduced again as follows:

[0199] The obtained first optimal matching point pairs are topological nodes that need to satisfy having neighbors and the same category. Here, it is necessary to process topological nodes without neighbors, and the steps are as follows:

[0200] (1) Obtain the solved target rotation transformation and target translation transformation;

[0201] (2) Traverse all topological nodes A in the positioning topological interval;

[0202] a. Apply the target rotation transformation and target translation transformation to the traversed topological nodes;

[0203] b. Set a second preset distance threshold F, such as 1.5. This second preset distance threshold should be larger than the threshold for finding neighbor nodes before;

[0204] c. Traverse all topological nodes B in the current mapping topological interval;

[0205] ① If the topological node A and the topological node B are of different categories, skip the current topological node in the current mapping topological interval being traversed;

[0206] ② If the distance between the topological node A and the topological node B exceeds the second preset distance threshold F, skip the current topological node in the current mapping topological interval being traversed;

[0207] ③ If either the topological node A or the topological node B is a point with neighbors, skip the current topological node in the current mapping topological interval being traversed;

[0208] ④ Finally, the non - skipped A and B are the second - best matching point pairs, that is, topological nodes without neighbors but can be matched.

[0209] Figure 7 The first - best matching point pairs and the second - best matching point pairs are shown. Among them, the two frames respectively represent two topological intervals 1 and 2, the red points represent columns, the green points represent parking space corner points, and the straight lines connecting the topological nodes in the two topological intervals indicate that they are the best matching point pairs found through the above process.

[0210] (17) Calculate the matching success ratio based on the first - best matching point pairs and the second - best matching point pairs;

[0211] Specifically, the matching success ratio = (the sum of the first - best matching point pairs and the second - best matching point pairs) divided by the total number of topological point pairs.

[0212] (18) If the matching success ratio is greater than the first preset ratio threshold, use the current mapping topological interval as a candidate mapping topological interval for the positioning topological interval.

[0213] Specifically, the obtained multiple matching success ratios can also be sorted in descending order, the first few matching success ratios are extracted, and the corresponding mapping topological intervals are used as candidate mapping topological intervals for the positioning topological interval.

[0214] In an alternative embodiment of the present invention, each shape context descriptor corresponding to the positioning topological interval is matched with each shape context descriptor corresponding to each candidate mapping topological interval, and the target matching result of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate mapping topological interval is determined according to the second matching result, which specifically includes the following steps:

[0215] (1) Calculate the similarity matrix of each shape context descriptor corresponding to the positioning topological interval and each shape context descriptor corresponding to each candidate mapping topological interval by using the chi-square formula;

[0216] Specifically, one omnidirectional semantic segmentation map corresponds to one shape context descriptor. Therefore, the similarity matrix of the shape context descriptors here is the similarity matrix of the shape context descriptor of a target omnidirectional semantic segmentation map in the positioning topological interval and the shape context descriptor of an omnidirectional semantic segmentation map in the candidate mapping topological interval, representing the similarity between each point in the downsampled contour points in the target omnidirectional semantic segmentation map and each point in the downsampled contour points in the omnidirectional semantic segmentation map.

[0217] (2) Process each similarity matrix by using the Hungarian matching strategy to obtain multiple matching point pair results;

[0218] Specifically, as Figure 8 shown, it is a matching point pair result of the points in the downsampled contour points in a target omnidirectional semantic segmentation map and the points in the downsampled contour points in an omnidirectional semantic segmentation map, which is the result displayed after splicing the two graphs of the downsampled contour points. In this way, multiple matching point pair results can be obtained.

[0219] (3) Determine the matching of the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate mapping topological interval according to the target matching point pair result with the largest number of matches in the multiple matching point pair results, where the first target shape context descriptor and the second target shape context descriptor are the shape context descriptors corresponding to the target matching point pair result.

[0220] As can be seen from the above description, the shape context descriptor is used to extract descriptors from a single-frame image and is also matched with a single-frame image. The matching between points is at the centimeter level, and the pose accuracy solved by using these matching points is relatively high. However, if the trigger frequency is too high, it will result in a relatively high requirement for computing power in practical applications.

[0221] The range of the random walk descriptor depends on the range of the generated topological interval. Because unique segmentation categories such as columns need to be used, the topological interval is at least more than 10 meters. For the matching of descriptors in this range, the pose solved sometimes has insufficient accuracy.

[0222] Therefore, a strategy of fusing two descriptors is designed. A topological interval is generated every relatively large distance, such as 30 meters. In each topological interval, a shape context descriptor is saved every few meters, such as 6 meters.

[0223] The method of the present invention combines two different types of feature descriptors. First, the shape context descriptor is extracted through the panoramic semantic segmentation map to capture local geometric features. Second, the column and parking space corner points, arrows, speed bumps and other categories with relatively high and stable occurrence frequencies (i.e., the target clustering center points) are extracted through the panoramic semantic segmentation map to construct a topological node graph. Then, the random walk descriptor is extracted based on the topological node graph to represent the global structure information. Finally, the two descriptors are used in combination. The random walk descriptor is used for rough matching to lock the topological interval of the similar scene, and then the shape context descriptor is used to further refine the range of the similar scene, so as to enhance the robustness of scene recognition.

[0224] The method of the present invention comprehensively utilizes the local geometric features of the context descriptor and the global structure information of the random walk descriptor. First, due to the characteristic of the large range of the random walk descriptor, rough matching is performed to obtain several candidate topological intervals, and then the shape context descriptors within the topological intervals are finely matched. Such a fusion strategy not only reduces the frequency of matching of the shape context descriptor and reduces the consumption of computing power, but also the two descriptors form a cross-validation effect, making the result of scene recognition more accurate and robust.

[0225] Embodiment 2:

[0226] The embodiment of the present invention also provides a scene encoding and encoding matching device for an underground parking lot. The scene encoding and encoding matching device for the underground parking lot is mainly used to execute the scene encoding and encoding matching method for the underground parking lot provided in Embodiment 1 of the present invention. The following is a specific introduction to the scene encoding and encoding matching device for the underground parking lot provided by the embodiment of the present invention.

[0227] Figure 9 is a schematic diagram of a scene encoding and encoding matching device for an underground parking lot according to an embodiment of the present invention. As Figure 9 shown, the device mainly includes: an acquisition and construction unit 10, a random walk descriptor extraction unit 20, a shape context descriptor extraction unit 30, a first matching unit 40, and a second matching unit 50, wherein:

[0228] An acquisition and construction unit is used to acquire a multi-frame surround view semantic segmentation map within a topological interval, and construct a topological node map corresponding to the topological interval according to the target cluster center point extracted from the multi-frame surround view semantic segmentation map, wherein the topological interval is a preset distance range, and the topological interval includes: multiple mapping topological intervals and one positioning topological interval;

[0229] A random walk descriptor extraction unit, used for extracting a random walk descriptor of a topological node graph according to the topological node graph;

[0230] A shape context descriptor extraction unit is used to extract a preset number of target surround view semantic segmentation maps at equal intervals from the multi-frame surround view semantic segmentation maps within the topological interval, and extract corresponding shape context descriptors according to the target surround view semantic segmentation maps;

[0231] A first matching unit is used to match the random walk descriptor corresponding to the positioning topology interval with the random walk descriptor corresponding to each mapping topology interval, and determine a candidate mapping topology interval matching the positioning topology interval in the mapping topology interval according to the first matching result;

[0232] The second matching unit is used to match each shape context descriptor corresponding to the positioning topology interval with each shape context descriptor corresponding to each candidate mapping topology interval, and determine the target matching result of the first target shape context descriptor of the positioning topology interval and the second target shape context descriptor of the candidate mapping topology interval according to the second matching result.

[0233] In an embodiment of the present invention, a scene encoding and encoding matching device for an underground parking lot is provided, including: obtaining multiple frames of panoramic semantic segmentation maps within a topological interval, and constructing a topological node map corresponding to the topological interval according to the target clustering center points extracted from the multiple frames of panoramic semantic segmentation maps, where the topological interval is a preset distance range, and the topological interval includes: a plurality of mapping topological intervals and a positioning topological interval; extracting a random walk descriptor of the topological node map according to the topological node map; equally spacing and extracting a preset number of target panoramic semantic segmentation maps from the multiple frames of panoramic semantic segmentation maps within the topological interval, and extracting corresponding shape context descriptors according to the target panoramic semantic segmentation maps; matching the random walk descriptor corresponding to the positioning topological interval with the random walk descriptors corresponding to each mapping topological interval respectively, and determining a candidate mapping topological interval that matches the positioning topological interval in the mapping topological intervals according to the first matching result; matching each shape context descriptor corresponding to the positioning topological interval with each shape context descriptor corresponding to each candidate mapping topological interval respectively, and determining a target matching result that the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate mapping topological interval according to the second matching result. It can be seen from the above description that in the scene encoding and encoding matching device for the underground parking lot of the present invention, first, a candidate mapping topological interval that matches the positioning topological interval is determined in the mapping topological intervals through the random walk descriptor, and then a target matching result that the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate mapping topological interval is determined through the shape context descriptor, that is, first, a rough match is performed through the random walk descriptor to determine a candidate mapping topological interval that matches the positioning topological interval, and then a fine match is performed through the shape context descriptor to finally obtain a target matching result that the first target shape context descriptor of the positioning topological interval matches the second target shape context descriptor of the candidate mapping topological interval. The above process of rough matching greatly reduces the matching workload of subsequent fine matching, improves the matching efficiency of the overall matching, and the above process of fine matching has good accuracy, and the matching between points is centimeter-level. By fusing the above two matching methods, a cross-verification effect is also formed, making the result of scene recognition more accurate and robust, and alleviating the technical problem that in the traditional scene recognition process, it is impossible to take into account both the efficiency (i.e., real-time performance) and accuracy of scene matching.

[0234] Optionally, the acquisition and construction unit is further configured to: within the topological interval, acquire the parking lot images collected by the vehicle-mounted surround-view camera, and perform projection stitching on the acquired parking lot images each time to obtain multiple frames of bird's-eye view images within the topological interval; perform semantic segmentation on each frame of bird's-eye view image to obtain multiple frames of surround-view semantic segmentation images within the topological interval, where the categories of semantic segmentation include: parking lines, solid lines, dashed lines, columns, speed bumps; extract the target clustering center points from the multiple frames of surround-view semantic segmentation images, where the target clustering center points include: column clustering center points, parking space corner point clustering center points, arrow clustering center points, speed bump clustering center points; convert the target clustering center points to the world coordinate system to obtain topological nodes, and further obtain a topological node graph composed of topological nodes.

[0235] Optionally, the random walk descriptor extraction unit is further configured to: determine non-independent points and independent points among the topological nodes in the topological node graph according to a first preset distance threshold, where a non-independent point is a topological node in which there are other topological nodes within the first preset distance threshold from itself; determine the dimension of the random walk descriptor of the non-independent points, where the dimension is K*1 dimension, K is the nth power of m, m represents the number of categories of all non-independent points in the topological node graph, and n represents the step length of the random walk; traverse the non-independent points, and determine each neighbor point of the current non-independent point among the topological nodes in the topological node graph according to the step length of the random walk and the first preset distance threshold, where the neighbor points include: direct neighbor points and / or indirect neighbor points, a direct neighbor point is a topological node in the topological node graph whose distance from the current non-independent point is less than the first preset distance threshold, and when the indirect neighbor point is a multi-level indirect neighbor point, each level of indirect neighbor point is a topological node in the topological node graph whose distance from its upper-level indirect neighbor point is less than the first preset distance threshold, and when the indirect neighbor point is a first-level indirect neighbor point, its upper-level indirect neighbor point is a direct neighbor point; calculate the index of the category combination according to the index calculation formula: the category of the current non-independent point * m to the power of n-1 + the category of the direct neighbor point * m to the power of n-2 + the category of the first-level indirect neighbor point * m to the power of n-3 + the category of the second-level indirect neighbor point * m to the power of n-4 +... + the category of the last-level indirect neighbor point * m to the power of n-n, where the category combination is: the category of the current non-independent point, the category of the direct neighbor point, the category of the first-level indirect neighbor point, the category of the second-level indirect neighbor point,..., the category of the last-level indirect neighbor point; add 1 to the element at the position of the random walk descriptor corresponding to the index, and further obtain the random walk descriptor of the current non-independent point, and obtain the non-independent point random walk descriptor according to the random walk descriptors of all non-independent points; obtain the random walk descriptor of the independent points, and determine the random walk descriptor of the topological node graph according to the non-independent point random walk descriptor and the random walk descriptor of the independent points, where the dimension of the random walk descriptor of the independent points is the same as the dimension of the random walk descriptor of the non-independent points.

[0236] Optionally, the shape context descriptor extraction unit is further configured to: extract the contour points of the target line with the category of line in the target panoramic semantic segmentation map; downsample the contour points to obtain the downsampled contour points; for each downsampled contour point, construct a polar logarithmic coordinate system centered thereon, wherein the polar logarithmic coordinate system includes: P angles and Q distances, thereby obtaining P*Q grid regions; map the other downsampled contour points around the central downsampled contour point into the P*Q grid regions, and count the number of downsampled contour points in each grid region; perform normalization processing on the number of downsampled contour points in each grid region to obtain the normalized number of downsampled contour points in each grid region; generate a shape context descriptor of the central downsampled contour point according to the normalized number, and determine the shape context descriptor corresponding to the target panoramic semantic segmentation map according to the shape context descriptors of all the central downsampled contour points.

[0237] Optionally, the first matching unit is further configured to: construct a scoring matrix, where the number of rows of the scoring matrix is the number of topological nodes in the positioning topological interval, the number of columns of the scoring matrix is the number of topological nodes in the current mapping topological interval, and the values of all elements in the scoring matrix are 0; traverse the random walk descriptors of each row in the positioning topological interval; if the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval is an independent point, skip the random walk descriptor of the current row in the positioning topological interval; if the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval is a non-independent point, traverse the random walk descriptors of each row in the current mapping topological interval; if the category of the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval is different from the category of the topological node corresponding to the random walk descriptor of the current row in the current mapping topological interval, skip the random walk descriptor of the current row in the current mapping topological interval; if the category of the topological node corresponding to the random walk descriptor of the current row in the positioning topological interval is the same as the category of the topological node corresponding to the random walk descriptor of the current row in the current mapping topological interval, calculate the matching similarity between the random walk descriptor of the current row in the positioning topological interval and the random walk descriptor of the current row in the current mapping topological interval; determine whether the random walk descriptor of the current row in the positioning topological interval and the random walk descriptor of the current row in the current mapping topological interval are proportional; if they are proportional, set the matching similarity to 0 and fill the matching similarity of 0 into the position in the scoring matrix corresponding to the current row of the positioning topological interval and the current row of the current mapping topological interval; if they are not proportional, fill the matching similarity into the position in the scoring matrix corresponding to the current row of the positioning topological interval and the current row of the current mapping topological interval until the traversal of the positioning topological interval and the current mapping topological interval is completed, thereby obtaining a target scoring matrix, where each element in the target scoring matrix represents the similarity between the topological node in the positioning topological interval corresponding to its corresponding row and the topological node in the current mapping topological interval corresponding to its corresponding column; determine the first optimal matching point pair and the corresponding target rotation transformation and target translation transformation according to the target scoring matrix; traverse the topological nodes in the positioning topological interval, and transform the current topological node in the positioning topological interval by using the target rotation transformation and target translation transformation to obtain the transformed current topological node; traverse the topological nodes in the current mapping topological interval; if the category of the current topological node in the positioning topological interval is different from the category of the current topological node in the current mapping topological interval, skip the current topological node in the current mapping topological interval; if the distance between the transformed current topological node and the current topological node in the current mapping topological interval is greater than a second preset distance threshold, skip the current topological node in the current mapping topological interval; if the current topological node in the positioning topological interval is a non-independent point, or, the current topological node in the current mapping topological interval is a non-independent point, skip the current topological node in the current mapping topological interval;Take the current topological nodes of the positioning topological interval that have not been skipped and the current topological nodes of the current mapping topological interval as the second optimal matching point pairs; calculate the matching success ratio based on the first optimal matching point pairs and the second optimal matching point pairs; if the matching success ratio is greater than the first preset ratio threshold, then take the current mapping topological interval as the candidate mapping topological interval of the positioning topological interval.

[0238] Optionally, the first matching unit is further configured to: determine target elements in the target scoring matrix whose element values are greater than the preset similarity threshold, and determine candidate matching point pairs according to the target elements; randomly select a target number of candidate matching point pairs from the candidate matching point pairs, and take the topological nodes belonging to the positioning topological interval in the target number of candidate matching point pairs as target points, and take the topological nodes belonging to the current mapping topological interval in the target number of candidate matching point pairs as source points; calculate the rotation transformation and translation transformation between the source points and the target points according to the target number of candidate matching point pairs; transform each target point by using the rotation transformation and translation transformation to obtain the transformed target points; determine whether the transformed target points are included within the third preset distance threshold of each source point; if so, increment the number of matching points by 1 to obtain the target number of matching points; if the target number of matching points is less than the preset number threshold, determine that the target number of candidate matching point pairs are incorrect matching point pairs, and return to execute the process of randomly selecting a target number of candidate matching point pairs from the candidate matching point pairs until the target number of matching points is greater than the preset number threshold; if the target number of matching points is greater than the preset number threshold, determine that the target number of candidate matching point pairs are the first optimal matching point pairs, and take the corresponding rotation transformation and translation transformation as the target rotation transformation and target translation transformation.

[0239] Optionally, the second matching unit is further configured to: calculate the similarity matrix of each shape context descriptor corresponding to the positioning topological interval and each shape context descriptor corresponding to each candidate mapping topological interval by using the chi-square formula; process each similarity matrix by using the Hungarian matching strategy to obtain multiple matching point pair results; determine that the first target shape context descriptor of the positioning topological interval and the second target shape context descriptor of the candidate mapping topological interval are matched according to the target matching point pair result with the largest number of matches among the multiple matching point pair results, where the first target shape context descriptor and the second target shape context descriptor are the shape context descriptors corresponding to the target matching point pair result.

[0240] The device provided by the embodiments of the present invention has the same implementation principle and the same technical effects as those of the foregoing method embodiments. For brief description, for the parts not mentioned in the device embodiments, reference may be made to the corresponding content in the foregoing method embodiments.

[0241] Such as Figure 10As shown in the figure, an electronic device 600 provided by an embodiment of the present application includes: a processor 601, a memory 602, and a bus. The memory 602 stores machine-readable instructions executable by the processor 601. When the electronic device runs, the processor 601 communicates with the memory 602 through the bus. The processor 601 executes the machine-readable instructions to perform the steps of the above-described scenario encoding and encoding matching method for an underground parking lot.

[0242] Specifically, the above-mentioned memory 602 and processor 601 can be general-purpose memory and processor, and no specific limitation is made here. When the processor 601 runs the computer program stored in the memory 602, it can execute the above-described scenario encoding and encoding matching method for an underground parking lot.

[0243] The processor 601 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor 601 or instructions in the form of software. The above-mentioned processor 601 can be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it can also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field-programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. This storage medium is located in the memory 602, and the processor 601 reads the information in the memory 602 and combines its hardware to complete the steps of the above method.

[0244] Corresponding to the above-mentioned scenario encoding and encoding matching method for the underground parking lot, an embodiment of the present application further provides a computer-readable storage medium. The computer-readable storage medium stores machine-executable instructions. When the computer-executable instructions are called and run by a processor, the computer-executable instructions cause the processor to run the steps of the above-mentioned scenario encoding and encoding matching method for the underground parking lot.

[0245] The scenario encoding and encoding matching device for the underground parking lot provided by the embodiments of the present application may be specific hardware on a device or software or firmware installed on the device. For the device provided by the embodiments of the present application, the implementation principle and the technical effects produced are the same as those of the foregoing method embodiments. For the sake of brief description, for the parts not mentioned in the device embodiments, reference may be made to the corresponding content in the foregoing method embodiments. Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the foregoing-described systems, devices, and units can all refer to the corresponding processes in the above method embodiments, and will not be repeated here.

[0246] In the embodiments provided by 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 illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces. The indirect coupling or communication connection of the devices or units may be in an electrical, mechanical, or other form.

[0247] For another example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of devices, methods, and computer program products according to multiple embodiments of the present application. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0248] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0249] In addition, each functional unit in the embodiments provided in this application may be integrated into a processing unit, may exist physically separately for each unit, or two or more units may be integrated into one unit.

[0250] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable an electronic device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the scenario encoding and encoding matching method for the underground parking lot described in each embodiment of this application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0251] It should be noted that: similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In addition, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0252] Finally, it should be noted that: the above-described embodiments are only specific implementation manners of this application, used to illustrate the technical solutions of this application, rather than limiting it. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: any person skilled in the art within the technical scope disclosed in this application can still modify the technical solutions recorded in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements for some of the technical features; and these modifications, changes, or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A scene coding and coding matching method for an underground parking lot, characterized in that: include: Acquire a multi-frame surround view semantic segmentation map within a topological interval, and construct a topological node map corresponding to the topological interval according to the target cluster center points extracted from the multi-frame surround view semantic segmentation map, wherein the topological interval is a preset distance range, and the topological interval includes: a plurality of mapping topological intervals and a positioning topological interval; Extracting a random walk descriptor of the topological node graph according to the topological node graph; Extracting a preset number of target surround view semantic segmentation maps at equal intervals from the multi-frame surround view semantic segmentation maps within the topological interval, and extracting corresponding shape context descriptors according to the target surround view semantic segmentation maps; Matching the random walk descriptor corresponding to the positioning topology interval with the random walk descriptor corresponding to each of the mapping topology intervals respectively, and determining a candidate mapping topology interval matching the positioning topology interval in the mapping topology interval according to the first matching result; Each shape context descriptor corresponding to the positioning topology interval is matched with each shape context descriptor corresponding to each of the candidate mapping topology intervals, and a target matching result of matching the first target shape context descriptor of the positioning topology interval with the second target shape context descriptor of the candidate mapping topology interval is determined according to the second matching result.

2. The method according to claim 1, characterized in that Get the multi-frame surround semantic segmentation map within the topological interval, including: In the topological interval, a parking lot image captured by a vehicle-mounted surround view camera is acquired, and the parking lot image acquired each time is projected and spliced ​​to obtain a multi-frame bird's-eye view in the topological interval; Performing semantic segmentation on each frame of the bird's-eye view to obtain a multi-frame surround semantic segmentation map within the topological interval, wherein the categories of the semantic segmentation include: stop line, solid line, dotted line, pillar, speed bump; Constructing a topological node graph corresponding to the topological interval according to the target cluster center points extracted from the multiple frames of the surround semantic segmentation graph, including: Extracting the target cluster center points from the multiple frames of the surround semantic segmentation map, wherein the target cluster center points include: pillar cluster center points, parking space corner cluster center points, arrow cluster center points, and speed bump cluster center points; The target cluster center point is converted to the world coordinate system to obtain a topological node, and then a topological node graph composed of the topological nodes is obtained.

3. The method according to claim 1, characterized in that Extracting a random walk descriptor of the topological node graph according to the topological node graph includes: Determine a non-independent point and an independent point in the topological nodes in the topological node graph according to a first preset distance threshold, wherein the non-independent point is a topological node in the topological node that has other topological nodes within the first preset distance threshold from itself; Determine the dimension of the random walk descriptor of the non-independent point, wherein the dimension is K*1, K is m to the power of n, m represents the number of categories of all non-independent points in the topological node graph, and n represents the step length of the random walk; Traversing the non-independent points, determining each neighbor point of the current non-independent point in the topological nodes in the topological node graph according to the step length of the random walk and the first preset distance threshold, wherein the neighbor points include: direct neighbor points and / or indirect neighbor points, the direct neighbor point is a topological node in the topological node graph whose distance from the current non-independent point is less than the first preset distance threshold, when the indirect neighbor point is a multi-level indirect neighbor point, each level of indirect neighbor point is a topological node in the topological node graph whose distance from its previous level of indirect neighbor point is less than the first preset distance threshold, when the indirect neighbor point is a first-level indirect neighbor point, its previous level of indirect neighbor point is the direct neighbor point; Calculate the index of the category combination according to the index calculation formula: the category of the current non-independent point*m to the n-1 power + the category of the direct neighbor point*m to the n-2 power + the category of the first-level indirect neighbor point*m to the n-3 power + the category of the second-level indirect neighbor point*m to the n-4 power +…+ the category of the last-level indirect neighbor point*m to the nn power, wherein the category combination is: the category of the current non-independent point, the category of the direct neighbor point, the category of the first-level indirect neighbor point, the category of the second-level indirect neighbor point,…, the category of the last-level indirect neighbor point; Add 1 to the element at the position of the random walk descriptor corresponding to the index, thereby obtaining the random walk descriptor of the current non-independent point, and obtain the random walk descriptor of the non-independent point based on the random walk descriptors of all non-independent points; A random walk descriptor of an independent point is obtained, and a random walk descriptor of the topological node graph is determined according to the random walk descriptor of the non-independent point and the random walk descriptor of the independent point, wherein the dimension of the random walk descriptor of the independent point is the same as the dimension of the random walk descriptor of the non-independent point.

4. The method according to claim 1, characterized in that: Extracting a corresponding shape context descriptor according to the target surround semantic segmentation map includes: Extracting contour points of target lines of line category in the target surround semantic segmentation map; Downsampling the contour points to obtain downsampled contour points; For each of the downsampled contour points, a polar logarithmic coordinate system centered on the downsampled contour point is constructed, wherein the polar logarithmic coordinate system includes: P angles and Q distances, thereby obtaining P*Q grid areas; Mapping other downsampled contour points around the central downsampled contour point to the P*Q grid areas, and counting the number of downsampled contour points in each of the grid areas; Normalizing the number of downsampled contour points in each of the grid areas to obtain a normalized number of downsampled contour points in each of the grid areas; A shape context descriptor of the downsampled contour points of the center is generated according to the normalized quantity, and a shape context descriptor corresponding to the target surround semantic segmentation map is determined according to the shape context descriptors of all the downsampled contour points of the center.

5. The method according to claim 1, characterized in that: Matching the random walk descriptor corresponding to the positioning topology interval with the random walk descriptor corresponding to each of the mapping topology intervals respectively, and determining a candidate mapping topology interval matching the positioning topology interval in the mapping topology interval according to a first matching result, including: Constructing a scoring matrix, wherein the number of rows of the scoring matrix is ​​the number of topological nodes in the positioning topological interval, the number of columns of the scoring matrix is ​​the number of topological nodes in the current mapping topological interval, and the value of each element in the scoring matrix is ​​0; Traversing the random walk descriptor of each row of the positioning topology interval; If the topological node corresponding to the random walk descriptor of the current row of the positioning topological interval is an independent point, skipping the random walk descriptor of the current row of the positioning topological interval; If the topological node corresponding to the random walk descriptor of the current row of the positioning topological interval is a non-independent point, traverse the random walk descriptors of each row of the current mapping topological interval; If the category of the topological node corresponding to the random walk descriptor of the current row of the positioning topology interval is different from the category of the topological node corresponding to the random walk descriptor of the current row of the current mapping topology interval, skip the random walk descriptor of the current row of the current mapping topology interval; If the topological node corresponding to the random walk descriptor of the current row of the positioning topology interval is of the same category as the topological node corresponding to the random walk descriptor of the current row of the current mapping topology interval, then the matching similarity between the random walk descriptor of the current row of the positioning topology interval and the random walk descriptor of the current row of the current mapping topology interval is calculated; Determine whether the random walk descriptor of the current row of the positioning topology interval is proportional to the random walk descriptor of the current row of the current mapping topology interval; If they are proportional, the matching similarity is set to 0, and the matching similarity of 0 is filled into the position corresponding to the current row of the positioning topology interval and the current row of the current mapping topology interval in the scoring matrix; If they are not proportional, the matching similarity is filled into the position corresponding to the current row of the positioning topology interval and the current row of the current mapping topology interval in the scoring matrix until the traversal of the positioning topology interval and the current mapping topology interval is completed, thereby obtaining a target scoring matrix, wherein each element in the target scoring matrix represents the similarity between the topological node of the positioning topology interval corresponding to its corresponding row and the topological node of the current mapping topology interval corresponding to its corresponding column; Determine a first optimal matching point pair and a corresponding target rotation transformation and target translation transformation according to the target scoring matrix; Traversing the topological nodes of the positioning topological interval, transforming the current topological node of the positioning topological interval by using the target rotation transformation and the target translation transformation to obtain the transformed current topological node; Traversing the topological nodes of the current mapping topological interval; If the current topological node of the positioning topological interval is of a different category from the current topological node of the current mapping topological interval, skipping the current topological node of the current mapping topological interval; If the distance between the transformed current topological node and the current topological node of the current mapping topological interval is greater than a second preset distance threshold, skipping the current topological node of the current mapping topological interval; If the current topological node of the positioning topological interval is a non-independent point, or the current topological node of the current mapping topological interval is a non-independent point, then skipping the current topological node of the current mapping topological interval; Taking the current topological node of the positioning topological interval that has not been skipped and the current topological node of the current mapping topological interval as the second optimal matching point pair; Calculating a matching success ratio based on the first optimal matching point pair and the second optimal matching point pair; If the matching success ratio is greater than a first preset ratio threshold, the current mapping topology interval is used as a candidate mapping topology interval for the positioning topology interval.

6. The method according to claim 5, characterized in that Determining a first optimal matching point pair and a corresponding target rotation transformation and target translation transformation according to the target scoring matrix includes: Determine a target element whose element value is greater than a preset similarity threshold in the target scoring matrix, and determine a candidate matching point pair according to the target element; Randomly select a target number of candidate matching point pairs from the candidate matching point pairs, and use the topological nodes belonging to the positioning topological interval among the target number of candidate matching point pairs as target points, and use the topological nodes belonging to the current mapping topological interval among the target number of candidate matching point pairs as source points; Calculate the rotation transformation and translation transformation between the source point and the target point according to the target number of candidate matching point pairs; Transforming each of the target points using the rotation transformation and the translation transformation to obtain a transformed target point; Determine whether the converted target point is included within a third preset distance threshold of each of the source points; If it is included, add 1 to the matching point to get the target matching point; If the target number of matching points is less than the preset number threshold, the target number of candidate matching point pairs are determined to be incorrect matching point pairs, and the process of randomly selecting the target number of candidate matching point pairs from the candidate matching point pairs is returned to execute until the target number of matching points is greater than the preset number threshold; If the target matching point number is greater than the preset number threshold, the target number of candidate matching point pairs are determined as the first optimal matching point pairs, and the corresponding rotation transformation and translation transformation are used as the target rotation transformation and the target translation transformation.

7. The method according to claim 1, characterized in that Matching each shape context descriptor corresponding to the positioning topology interval with each shape context descriptor corresponding to each of the candidate mapping topology intervals, and determining a target matching result of matching the first target shape context descriptor of the positioning topology interval with the second target shape context descriptor of the candidate mapping topology interval according to the second matching result, including: Calculate the similarity matrix of each shape context descriptor corresponding to the positioning topology interval and each shape context descriptor corresponding to each candidate mapping topology interval using the chi-square formula; The Hungarian matching strategy is used to process each similarity matrix to obtain multiple matching point pair results; The first target shape context descriptor of the positioning topology interval and the second target shape context descriptor of the candidate mapping topology interval are matched according to the target matching point pair result with the largest number of matches among the multiple matching point pair results, wherein the first target shape context descriptor and the second target shape context descriptor are shape context descriptors corresponding to the target matching point pair results.

8. A scene coding and coding matching device for an underground parking lot, characterized in that: include: An acquisition and construction unit, used to acquire a multi-frame surround view semantic segmentation map within a topological interval, and construct a topological node map corresponding to the topological interval according to the target cluster center points extracted from the multi-frame surround view semantic segmentation map, wherein the topological interval is a preset distance range, and the topological interval includes: a plurality of mapping topological intervals and a positioning topological interval; A random walk descriptor extraction unit, used for extracting a random walk descriptor of the topological node graph according to the topological node graph; A shape context descriptor extraction unit is used to extract a preset number of target surround view semantic segmentation maps at equal intervals from the multi-frame surround view semantic segmentation maps within the topological interval, and extract corresponding shape context descriptors according to the target surround view semantic segmentation maps; A first matching unit is used to match the random walk descriptor corresponding to the positioning topology interval with the random walk descriptor corresponding to each of the mapping topology intervals, and determine a candidate mapping topology interval matching the positioning topology interval in the mapping topology interval according to the first matching result; The second matching unit is used to match each shape context descriptor corresponding to the positioning topology interval with each shape context descriptor corresponding to each candidate mapping topology interval, and determine the target matching result of the first target shape context descriptor of the positioning topology interval and the second target shape context descriptor of the candidate mapping topology interval according to the second matching result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to execute the method according to any one of claims 1 to 7.

Images