Semantic feature descriptor determination method, apparatus, and electronic device

Abstract

The application provides a semantic feature descriptor determination method and device and electronic equipment, relates to the technical field of semantic descriptor application, and comprises the following steps: extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle; determining a category mapping corresponding to the target feature according to a semantic segmentation graph corresponding to the current frame image; and determining a semantic feature descriptor in the current frame image based on the category mapping, thereby relieving the technical problems of low robustness and difficulty in meeting mass production requirements in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semantic feature descriptor, and in particular to a semantic feature descriptor determination method and device and electronic equipment. BACKGROUND

[0002] Feature descriptor refers to information designed in a certain way, which has a description effect on the pixels around the feature. It is designed according to the principle that features with similar appearances should have similar descriptors, and can be represented by a vector.

[0003] At present, feature descriptors are roughly divided into two categories: hand-crafted descriptors based on traditional methods and training descriptors based on deep learning methods. Traditional hand-crafted feature descriptors such as BRIEF, SIFT and SUFT are currently quite mature. They well describe the gray relationship of the pixels around the feature through artificial design, and realize feature matching. However, different light, weather and other factors have a great impact on the robustness of these methods, which greatly limits their application in outdoor environments. Descriptors based on deep learning emerged in such a background. However, this kind of method needs a large number of training samples, and the computing power and real-time performance are difficult to meet the mass production demand. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a semantic feature descriptor determination method, device and electronic equipment, which can alleviate the technical problems of low robustness and difficulty in meeting mass production demand in the prior art.

[0005] In a first aspect, an embodiment provides a semantic feature descriptor determination method, which comprises:

[0006] extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle;

[0007] determining a class mapping corresponding to the target feature according to a semantic segmentation map corresponding to the current frame image;

[0008] determining a semantic feature descriptor in the current frame image based on the class mapping.

[0009] In an optional implementation, the semantic feature descriptor comprises a semantic class of the target feature, a neighboring spatial semantic class of the target feature and a pixel distance of the target feature to the neighboring space.

[0010] In an optional implementation, the step of extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle comprises:

[0011] determine the target feature to be extracted according to a target object type corresponding to the preset requirement; wherein the target feature comprises one or more of the following: point feature, line feature and surface feature;

[0012] extract the target feature from a current frame image collected by the vehicle.

[0013] In an optional embodiment, the step of determining the class mapping corresponding to the target feature according to the semantic segmentation map corresponding to the current frame image comprises:

[0014] determine a second position of the target feature in the semantic segmentation map corresponding to the current frame image according to a first position of the target feature in the current frame image; wherein the current frame image and the semantic segmentation map corresponding to the current frame image have the same size parameter;

[0015] determine a semantic class corresponding to the target feature and a neighboring spatial semantic class of the target feature based on the second position in the semantic segmentation map.

[0016] In an optional embodiment, the method further comprises:

[0017] calculate the similarity between the semantic feature descriptor and the target descriptor in the current frame image; wherein the target descriptor is from a previous frame image and / or a preset map.

[0018] In an optional embodiment, the step of calculating the similarity between the semantic feature descriptor and the target descriptor in the current frame image comprises:

[0019] determine a weight coefficient according to the matching result of each semantic class in the semantic feature descriptor and the target descriptor in the current frame image;

[0020] determine the similarity between the semantic feature descriptor and the target descriptor based on the weight coefficient and the pixel distance of each target feature to the neighboring space.

[0021] In an optional embodiment, the method further comprises:

[0022] verify the actual application model of the semantic feature descriptor based on the similarity.

[0023] In a second aspect, an embodiment provides a device for determining a semantic feature descriptor, the device comprising:

[0024] an extraction module configured to extract a target feature meeting a preset requirement from a current frame image collected by a vehicle;

[0025] a first determination module configured to determine a class mapping corresponding to the target feature according to a semantic segmentation map corresponding to the current frame image;

[0026] A second determining module, based on the category mapping, determines semantic feature descriptors in the current frame image.

[0027] In a third aspect, embodiments provide an electronic device, comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor implements steps of the method according to any one of the preceding embodiments when executing the computer program.

[0028] In a fourth aspect, embodiments provide a machine readable storage medium, which stores machine executable instructions, and the machine executable instructions, when invoked and executed by a processor, cause the processor to implement steps of the method according to any one of the preceding embodiments.

[0029] The method, device and electronic device for determining semantic feature descriptors provided by the embodiments of the present application extract target features corresponding to a preset requirement from image frames collected by a vehicle, and obtain category mappings corresponding to each target feature from a semantic segmentation image corresponding to the image frames. Based on the category mappings, semantic feature descriptors corresponding to each target feature in the image frames can be determined. The feature descriptors determined in this way can guarantee feature description capability and can also be free from dependence on external factors such as light and weather.

[0030] Other features and advantages of the present disclosure will be described in the following description, or can be learned from the description, or can be determined without any doubt, or can be known by implementing the above-mentioned technologies of the present disclosure.

[0031] In order to make the above-mentioned purposes, features and advantages of the present disclosure more obvious and easy to understand, the following preferred embodiments are specifically described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without any creative labor.

[0033] Figure 1 A flow chart of a method for determining semantic feature descriptors provided by the embodiments of the present application;

[0034] Figure 2 A schematic diagram of a component structure of a semantic feature descriptor provided by the embodiments of the present application;

[0035] Figure 3 A semantic category order distribution diagram in a semantic feature descriptor provided for an embodiment of the present application;

[0036] Figure 4 A target feature diagram in a semantic segmentation map provided for an embodiment of the present application;

[0037] Figure 5 A functional module diagram of a determination device of a semantic feature descriptor provided for an embodiment of the present application;

[0038] Figure 6 A hardware architecture diagram of an electronic device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0039] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described below in detail with reference to the accompanying drawings. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0040] The current hand-crafted descriptor is greatly affected by the robustness of light, weather and other factors, and has great limitations in outdoor applications. On this basis, the descriptor can be trained based on a deep learning method. This kind of descriptor can get rid of the influence of outdoor environmental light, weather and other factors, but the acquisition of the descriptor requires a large number of training samples, and the cost requirement is high, and mass production is difficult to meet.

[0041] Based on this, the embodiments of the present application provide a determination method, device and electronic device of a semantic feature descriptor, which defines a more robust feature descriptor by using the stability of semantic information to light and weather, greatly improving the description ability of the feature.

[0042] To facilitate the understanding of the embodiments, first, a determination method of a semantic feature descriptor disclosed by the embodiments of the present application is described in detail. The method can be applied to a car machine, a vehicle controller, and can also be applied to an intelligent control device such as a host computer and a server.

[0043] Figure 1 A determination method flowchart of a semantic feature descriptor provided for an embodiment of the present application.

[0044] As shown in Figure 1 the method comprises the following steps:

[0045] Step S102, extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle.

[0046] The vehicle can collect the road on which the vehicle travels and the facilities around the road through a collection device such as a visual sensor, and obtain the image frame corresponding to each moment; different vehicles or application scenarios can have different preset requirements, for example, the vehicle involves a parking scenario, and the target feature can be related to lane lines, corner points and other features; the target object in the embodiment of the application can be understood as a static object, that is, the dynamic obstacle is an object that needs to be removed in the adjacent frame matching process or the map matching process. The static object includes buildings, traffic facilities, signboards, lamp poles, ground markings and the like, and the corner points, lines and surfaces on these static objects are target features to be studied and matched.

[0047] In step S104, the category mapping corresponding to the target feature is determined according to the semantic segmentation map corresponding to the current frame image.

[0048] Each image frame can obtain the corresponding semantic segmentation map through a deep learning network, the semantic segmentation map can be free from the influence of factors such as light and weather, and can identify the category of each object; the category mapping can be understood as a parameter that can know the category of the target feature itself and the category of each feature in the adjacent space of the target feature.

[0049] In step S106, the semantic feature descriptor in the current frame image is determined based on the category mapping.

[0050] Here, the semantic feature descriptor corresponding to the image frame can be defined according to the category mapping, and the descriptor can be free from the dependence on the influencing factors such as light and weather.

[0051] In the preferred embodiment of actual application, the target feature corresponding to the preset requirement is extracted from the image frame collected by the vehicle, the category mapping corresponding to each target feature is obtained from the semantic segmentation map corresponding to the image frame, and the semantic feature descriptor corresponding to each target feature in the image frame can be determined based on the category mapping. The feature descriptor determined in this way can not only guarantee the feature description ability, but also be free from the dependence on external factors such as light and weather.

[0052] In some embodiments, the semantic feature descriptor is composed of the semantic category of the target feature, the semantic category of the adjacent space of the target feature and the pixel distance from the target feature to the adjacent space.

[0053] As Figure 2As shown, the semantic feature descriptor includes SK, NK, and ND, wherein SK represents the semantic category of the target feature, the data type of which is unsigned char, NK represents the semantic category of the adjacent space of the target feature, the data type of which is unsigned char, and ND represents the pixel distance of the target feature to the adjacent space, the data type of which is unsigned int. NK and ND each correspond to 8 variables, that is, the adjacent space can be understood as the features adjacent to the target feature in 8 directions, i.e., up, down, left, right, top-left, bottom-left, top-right, and bottom-right; the pixel distance of the target feature to the adjacent space can be understood as the pixel distance from the target feature in 8 directions, i.e., up, down, left, right, top-left, bottom-left, top-right, and bottom-right, to the adjacent space feature.

[0054] It should be noted that the above 8 directions need to be pre-defined in order, and the specific pre-definition of the order can be adjusted according to actual conditions; the descriptor definition and the direction order definition need to be strictly one-to-one corresponding. That is Figure 2 The 17 variables corresponding to the descriptor in are in order, and 1-8 in the NK and ND variables are defined and arranged according to this order. Among them, Figure 3 As shown is a direction order definition diagram of a target feature, which Figure 3 If the order in changes, the generated descriptor needs to be adjusted accordingly.

[0055] In some embodiments, the target feature in the image frame is extracted according to the preset requirement, so as to facilitate the generation of the subsequent feature descriptor; here, the step S102 can be realized by the following steps, including:

[0056] Step 2.1), determining the target feature to be extracted according to the target object type corresponding to the preset requirement.

[0057] Among them, the target object type can be understood as the type of target object to be detected according to the preset requirement, and the actual target object type depends on the application scenario, and different types may need to extract different target features; the target feature includes one or more of the following: rectangular box feature, point feature, line feature, and surface feature. The point feature is the most common in image features, which is generally the corner point of an object or texture.

[0058] Step 2.2), extracting the target feature from the current frame image collected by the vehicle.

[0059] Among them, the target feature extraction of the image frame can be realized by using a conventional feature extractor.

[0060] In some embodiments, in order to obtain a descriptor with better feature description capability, the class mapping corresponding to the target feature can be obtained, and the step S104 can specifically include the following steps:

[0061] Step 3.1), determining a second position of the target feature in a semantic segmentation map corresponding to the current frame image according to the first position of the target feature in the current frame image.

[0062] Wherein, the current frame image and the semantic segmentation map corresponding to the current frame image have the same size and scale, so that the second position of the target feature in the semantic segmentation map can be determined according to the first position of the target feature in the image frame.

[0063] Step 3.2), determining a semantic class corresponding to the target feature and a neighboring spatial semantic class of the target feature based on the second position in the semantic segmentation map.

[0064] It should be noted that the semantic segmentation map is classified by deep learning for each pixel in the image frame, and a class label corresponding to each pixel is given, that is, based on the semantic segmentation map, it can be determined which pixel class is a building, which is a person, which is a tree, which is a lane line, etc. The semantic segmentation can be realized by using the prior art, and the semantic description sub is generated by using the semantic segmentation technology, which can describe a feature in a higher dimension. Therefore, the robustness of the positioning based on the feature descriptor is greatly improved. The traditional feature descriptor is easily affected by the environment and fails, resulting in a decrease in positioning ability or even loss of positioning. The semantic description sub provided in the embodiment of the present application has better applicability to the environment and stronger positioning robustness, and is not easy to lose positioning.

[0065] In order to make the embodiment of the present application more clearly understood, the target feature is taken as a point feature for example, as shown in Figure 4 The circle point is the target feature extracted from the image frame, and according to the semantic segmentation map, the target feature is determined to correspond to the class of building, and the neighboring spatial feature class in 8 directions is determined. Wherein, in the semantic segmentation map, each class will be defined as an ID, such as building 1, tree 2, sky 3, vehicle 4, etc. It should be noted that the image edge also needs to be defined as a class, that is, an independent ID is configured. The edge of each class can be understood as the boundary line of different filled color blocks in the semantic segmentation map, and the image edge refers to the upper, lower, left and right four boundaries of the semantic segmentation map, such as image edge IDs 101, 102, 103 and 104.

[0066] For example Figure 4The left side order 1 and the upper side order 3 of the target feature respectively extend to the edges (left boundary and upper boundary) of the semantic segmentation map, and the pixel distances of the target feature to the two edges are g and i respectively, and the other five directions respectively extend to the boundaries of other filled color blocks, that is, the categories of other filled color blocks are known and the pixel distances of the target feature to adjacent spatial features are known, such as the left upper order 2, the right upper order 4 and the right side order 5 respectively extending to the sky category, and the target feature respectively has the pixel distances a, b and c to them, the left lower order 8 and the lower side order 7 respectively extending to the tree category, and the target feature respectively has the pixel distances d and e to them, and the right lower order 6 extending to the vehicle category, and the target feature respectively has the pixel distance f to it.

[0067] On the basis of the foregoing embodiment, the semantic feature descriptor corresponding to the target feature can be as shown in Table 1:

[0068] Table 1

[0069]

[0070] In some embodiments, in order to measure the defined feature descriptor, the similarity of the feature descriptor is generally considered, and the method includes:

[0071] Step 4.1), calculating the similarity of the semantic feature descriptor in the current frame image and the target descriptor.

[0072] The target descriptor can be determined according to the application scenario thereof, and is from the previous frame image and / or a preset map. If the compared objects are two features in the previous and subsequent frame images, it is not determined whether they are the same point, and the descriptor is used to judge. For us here, it is A original image+A segmentation, and a feature is extracted. Taking a certain feature as an example, the feature is found in B original image+B segmentation which matches it, and the descriptor comparison is performed to find the correct feature association of the feature. Alternatively, the target descriptor corresponding to a certain point feature in the preset map is compared with the semantic feature descriptor, and if there is a difference, the preset range of the feature point is continuously searched. A feature point in a graph A will be given a priori according to the two frame image pose relationship of other sensors, and a theoretical corresponding pose is obtained in a graph B, and then all features in the preset range of the theoretical pose are matched to find the most similar point. If none of them is similar, the feature point in the graph A fails to be associated. The graph A includes a large number of features as described above, and the above process is a matching process of one of them. Each feature point needs to be operated once.

[0073] For example, weight coefficients are determined based on the matching results of semantic feature descriptors and target descriptors for each semantic category in the current frame image; the similarity between semantic feature descriptors and target descriptors is determined based on the weight coefficients and the pixel distance from each target feature to the adjacent space.

[0074] As an optional embodiment, the formula for calculating the semantic feature descriptor similarity is as follows:

[0075] S = A SK +B NK1 *C ND1 +B NK2 *C ND2 +B NK3 *C ND3 +B NK4 *C ND4 +B NK5 *C ND5

[0076] +B NK6 *C ND6 +B NK7 *C ND7 +B NK8 *C ND8

[0077] Where S is the distance between the target descriptor and the semantic feature descriptor. The smaller the value, the more similar the two are, and the higher the similarity.

[0078] For the weighting coefficient A SK In terms of values, if the SK values ​​of the two descriptors are not equal, that is, the target feature categories of the two descriptors are different, then A SK =100; if SK values ​​are equal, then A SK =0. Here, 100 is a default value, related to the segmentation image resolution and segmentation effect, and can be adjusted according to the actual situation. The larger this value, the greater the semantic weight of the feature itself. That is, if two feature points have different semantic categories, the larger this value, the larger the value of S, and the lower the similarity. Generally, the higher the resolution, the larger this value; the higher the classification confidence, the larger this value.

[0079] Regarding the value of the weight coefficient B, the eight B values ​​of the two descriptors are respectively... NK The comparison is performed sequentially, that is, the adjacent spatial categories in the direction corresponding to each target feature are compared in order; if the adjacent spatial categories in that direction are the same, the corresponding B... NK =1; Different corresponding B NK= 10. Here, 10 is a default value, which is related to the segmentation resolution and segmentation effect. The greater the value, the greater the semantic weight of the target feature neighborhood, that is, if the semantic categories of two neighborhoods are different, the greater the value, the greater the value of S, and the lower the similarity. Generally, the greater the resolution, the greater the value; the smaller the segmentation noise, the greater the value.

[0080] For the value of the weight coefficient C, C NK is the pixel distance of the target feature to the corresponding direction neighborhood feature according to the order, which is rounded down. For example, Figure 4 The pixel distance of the target feature on the right side of the order 5 to the corresponding direction neighborhood (adjacent space) feature of the sky category is 86 pixels, so the value of C ND5 is 86.

[0081] It should be noted that the weight coefficients A and B represent the respective loss residuals or penalties. The greater the penalty, the more different the two descriptors. Among them, the penalty weight of A is higher, because if the semantic categories of two feature points are different, they are probably not the same point. In order to ensure the reliability of the semantic segmentation result, the penalty weight, i.e., the weight coefficient, can be adjusted according to the actual ability of the segmentation network. If the missegmentation is less, the penalty weight can be further increased. B is similar, and the value is determined by judging whether the categories of the 8 direction neighborhoods are consistent. Details are not repeated here.

[0082] In some embodiments, the similarity can also be verified by applying the scene, such as the method further comprising:

[0083] Step 5.1), based on the similarity, checking the actual application model of the semantic feature descriptor.

[0084] If the similarity is high, the descriptor positioning accuracy of the current application scene is high, and if the similarity does not reach the threshold, the actual application model corresponding to the current application scene can be adjusted according to the similarity difference to ensure the reliability of subsequent scene application.

[0085] The embodiment of the present application proposes a hand-crafted descriptor combined with visual segmentation semantic information, which skillfully utilizes the stability of semantic information to light and weather, greatly improves the description ability of the feature, and does not need to be trained, and the calculation speed is obviously better than that of the deep learning descriptor.

[0086] As shown in Figure 5 , the embodiment of the present application further provides a determination device 200 of a semantic feature descriptor, which comprises:

[0087] The extraction module 201 extracts the target feature meeting the preset demand from the current frame image collected by the vehicle;

[0088] The first determining module 202 determines a class mapping corresponding to the target feature according to a semantic segmentation map corresponding to the current frame image.

[0089] The second determining module 203 determines a semantic feature descriptor in the current frame image based on the class mapping.

[0090] In some embodiments, the semantic feature descriptor includes a semantic class of the target feature, a neighboring spatial semantic class of the target feature, and a pixel distance of the target feature to a neighboring space.

[0091] In some embodiments, the extraction module 201 is further specifically configured to determine the target feature to be extracted according to a target object type corresponding to a preset requirement, wherein the target feature includes one or more of the following: point feature, line feature, and surface feature; and extract the target feature from a current frame image collected by a vehicle.

[0092] In some embodiments, the first determining module 202 is further specifically configured to determine a second position of the target feature in a semantic segmentation map corresponding to the current frame image according to a first position of the target feature in the current frame image, wherein the current frame image and the semantic segmentation map corresponding to the current frame image have the same size parameter; and determine a semantic class corresponding to the target feature and a neighboring spatial semantic class of the target feature based on the second position in the semantic segmentation map.

[0093] In some embodiments, the apparatus is further specifically configured to calculate a similarity between the semantic feature descriptor in the current frame image and a target descriptor, wherein the target descriptor is from a previous frame image and / or a preset map.

[0094] In some embodiments, the apparatus is further specifically configured to determine a weight coefficient according to a matching result of each semantic class in the semantic feature descriptor in the current frame image and the target descriptor; and determine the similarity between the semantic feature descriptor and the target descriptor based on the weight coefficient and a pixel distance of each target feature to a neighboring space.

[0095] In some embodiments, the apparatus is further specifically configured to verify an actual application model of the semantic feature descriptor based on the similarity.

[0096] Figure 6 A hardware architecture schematic diagram of the electronic device 300 is provided for the embodiments of the present application. Referring to FIG. 3, the electronic device 300 includes a processor 301, a memory 302, a communication interface 303, and a bus 304. Figure 6As shown, the electronic device 300 includes a machine readable storage medium 301 and a processor 302, and can further include a non-volatile storage medium 303, a communication interface 304 and a bus 305; wherein the machine readable storage medium 301, the processor 302, the non-volatile storage medium 303 and the communication interface 304 complete the communication among each other through the bus 305. The processor 302 can execute the above-described method for determining the semantic feature descriptor by reading and executing the determined machine executable instructions of the semantic feature descriptor in the machine readable storage medium 301.

[0097] The machine readable storage medium mentioned herein can be any electronic, magnetic, optical or other physical storage apparatus, and can contain or store information such as executable instructions, data, etc. For example, the machine readable storage medium can be a RAM (Random Access Memory), a volatile memory, a non-volatile memory, a flash memory, a storage drive (such as a hard disk drive), any type of storage disk (such as an optical disk, a DVD, etc.), or similar storage medium, or a combination thereof.

[0098] The non-volatile medium can be a non-volatile memory, a flash memory, a storage drive (such as a hard disk drive), any type of storage disk (such as an optical disk, a DVD, etc.), or similar non-volatile storage medium, or a combination thereof.

[0099] It can be understood that the specific operation method of each functional module in the embodiment can refer to the detailed description of the corresponding steps in the above method embodiment, which will not be repeated here.

[0100] The computer readable storage medium provided by the embodiment of the application stores a computer program, and the computer program code can implement the method for determining the semantic feature descriptor described in any of the above embodiments when executed, and the specific implementation can refer to the method embodiment, which will not be repeated here.

[0101] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and device can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0102] In addition, in the description of the embodiment of the application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integrally connected; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0103] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0104] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, and are not limiting. The protection scope of the present application is not limited thereto, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical range disclosed by the present application, or make equivalent replacement to some technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application.

Claims

1. A method for determining a semantic feature descriptor, characterized in that, The method comprises: extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle; determining a category mapping corresponding to the target feature according to a semantic segmentation map corresponding to the current frame image; determining a semantic feature descriptor in the current frame image based on the category mapping; the semantic feature descriptor comprises a semantic category of the target feature, a neighboring spatial semantic category of the target feature, and a pixel distance of the target feature to a neighboring space; the step of determining the category mapping corresponding to the target feature according to the semantic segmentation map corresponding to the current frame image comprises: determining a second position of the target feature in the semantic segmentation map corresponding to the current frame image according to a first position of the target feature in the current frame image; wherein the current frame image and the semantic segmentation map corresponding to the current frame image have the same size parameter; determining the semantic category corresponding to the target feature and the neighboring spatial semantic category of the target feature based on the second position in the semantic segmentation map.

2. The method of claim 1, wherein, The step of extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle comprises: determining a target feature to be extracted according to a target object type corresponding to the preset requirement; wherein the target feature comprises one or more of the following: point feature, line feature and surface feature; extracting the target feature from the current frame image collected by the vehicle.

3. The method of claim 1, wherein, The method further comprises: calculating a similarity between the semantic feature descriptor in the current frame image and a target descriptor; wherein the target descriptor is from a previous frame image and / or a preset map.

4. The method of claim 3, wherein, The step of calculating the similarity between the semantic feature descriptor in the current frame image and the target descriptor comprises: determining a weight coefficient according to a matching result of each semantic category in the semantic feature descriptor in the current frame image and the target descriptor; determining the similarity between the semantic feature descriptor and the target descriptor based on the weight coefficient and a pixel distance of each target feature to a neighboring space.

5. The method of claim 3, wherein, The method further comprises: verifying an actual application model of the semantic feature descriptor based on the similarity.

6. A device for determining a semantic feature descriptor, characterized in that The device comprises: an extraction module for extracting a target feature meeting a preset requirement from a current frame image collected by a vehicle; a first determination module for determining a category mapping corresponding to the target feature according to a semantic segmentation map corresponding to the current frame image; a second determination module for determining a semantic feature descriptor in the current frame image based on the category mapping; the semantic feature descriptor comprises a semantic category of the target feature, a neighboring spatial semantic category of the target feature, and a pixel distance of the target feature to a neighboring space; the first determination module is further configured to determine a second position of the target feature in the semantic segmentation map corresponding to the current frame image according to a first position of the target feature in the current frame image; wherein the current frame image and the semantic segmentation map corresponding to the current frame image have the same size parameter; and determine the semantic category corresponding to the target feature and the neighboring spatial semantic category of the target feature based on the second position in the semantic segmentation map.

7. An electronic device comprising a memory, a processor, the memory having stored therein a computer program executable on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the method of any one of claims 1 to 5.

8. A machine-readable storage medium, characterized in that, The machine readable storage medium stores machine executable instructions that, when called and executed by a processor, cause the processor to implement the steps of the method of any one of claims 1 to 5.

Images