A parking space detection model training method and electronic equipment

Abstract

This invention discloses a method for training a parking space detection model, comprising: identifying the type of parking space corner points in a parking space surround view image; if normal corner points exist in the parking space surround view image, then a normal corner point label is associated with the parking space surround view image; if difficult corner points exist in the parking space surround view image, then a difficult corner point label is associated with the parking space surround view image; training and correcting the parking space recognition model through a loss function; wherein, the normal corner point labels of the parking space surround view image are used as the first positive samples, and the loss value of the first positive samples directly participates in the training and correction process of the parking space recognition model; the difficult corner point labels of the parking space surround view image are used as the second positive samples. Compared with the prior art, the technical solution of this invention helps to reduce the negative impact of difficult targets on the model, while reducing the requirements for the judgment of the annotators, and improving the clarity and feasibility of the annotation standards. This invention also provides an electronic device that has the same beneficial effects.

Description

Technical Field

[0001] This invention relates to the field of automatic parking technology, and more particularly to a method for training a parking space detection model. This invention also relates to an electronic device. Background Technology

[0002] Parking space detection is one of the key perception tasks in automated parking systems. Locating parking spaces by detecting their corner points and directions is a novel method with advantages such as high versatility and low computational cost. Specifically, the corner point direction is detected through classification. The 360° angle is divided into K equal intervals, and each interval is treated as a binary classification problem to detect all directions of the corner point.

[0003] Deep learning algorithms applied to parking space detection typically employ supervised training, requiring manual annotation of training samples. This manual annotation relies on human understanding of annotation standards and subjective judgment of each sample. During the training phase, manually annotated locations are considered positive samples, while other background areas are considered negative samples. The goal of training is to improve the response values ​​of positive samples and suppress the response values ​​of negative samples. This is achieved by adjusting the model parameters through a loss function. The trained model can then be used for parking space detection tasks.

[0004] However, in automated parking scenarios, situations such as dim lighting, motion blur, and image stitching errors are unavoidable. These complex conditions can cause two negative impacts: First, they can affect the clarity and feasibility of annotation standards and place higher demands on the judgment of annotators, often leading to more contradictory annotation results, which is not conducive to model convergence. Second, they can generate more ambiguous and difficult targets. These targets contain both effective features for identifying the target and some spurious features. If they are treated as positive samples, spurious features will be introduced, leading to an increase in false detections. If they are treated as negative samples, some effective features will be suppressed, leading to an increase in false negatives.

[0005] How to reduce the negative impact of these difficult examples on the model and improve the clarity and feasibility of the annotation standards has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, the main objective of this invention is to provide a parking space detection model training method, as well as a parking space detection model training method and electronic device, which can reduce the negative impact of difficult examples on the model, while reducing the requirements for the judgment of annotation personnel and improving the clarity and feasibility of annotation standards.

[0007] To achieve the above objectives, this invention provides a method for training a parking space detection model based on parking space corner coordinate data, comprising the following steps:

[0008] Identify the type of parking space corner points in the parking space surround view image. The types of parking space corner points are divided into: normal corner points, difficult corner points, and non-corner points.

[0009] If there are normal corner points in the parking space surround view image, then a normal corner point label is associated with the parking space surround view image; and if there are difficult corner points in the parking space surround view image, then a difficult corner point label is associated with the parking space surround view image.

[0010] In the training of the deep learning-based parking space recognition model, the model is trained and corrected using a loss function.

[0011] The normal corner labels of the parking space surround view image are used as the first positive samples, and the loss value of the first positive samples directly participates in the training and correction process of the parking space recognition model.

[0012] The difficult corner labels of the parking space surround view image are used as the second positive samples. The loss value of the second positive sample is multiplied by the first adjustment coefficient δ1 and then participates in the training and correction process of the parking space recognition model. The first adjustment coefficient δ1 ∈ [0, 1).

[0013] It should be noted that when δ1 = 0, difficult examples no longer affect the model update;

[0014] Furthermore, in a more preferred embodiment of the present invention, the method further includes the steps of: identifying the labeled direction of the parking space corner point in the parking space surround view image; discretizing the continuous 0 to 360° into K angle intervals based on the parking space corner point, where K is a positive integer; marking the ±R angle intervals located to the left and right of the angle interval where the labeled direction is located as the neighboring direction; marking other angle intervals beyond the ±R angle intervals as other directions; in the training of the deep learning-based parking space recognition model, training and correcting the parking space recognition model through a loss function; enhancing the response value of the labeled direction in the process of participating in the training and correction of the parking space recognition model; suppressing the response values ​​of the neighboring direction and the other directions in the process of participating in the training and correction of the parking space recognition model, wherein the suppression effect of the response value of the neighboring direction is proportional to the product of the loss value of the neighboring direction in the training and correction process and the second control coefficient δ2, and δ2∈[0,1).

[0015] Furthermore, in a more preferred embodiment of the present invention, the step "suppressing the response value of the neighboring direction in the training and correction process of the parking space recognition model" specifically refers to: suppressing the response value of the neighboring direction in the training and correction process based on the deviation r between the neighboring direction and the labeled direction, where r∈[-R,R].

[0016] Furthermore, in a more preferred embodiment of the present invention, the step "suppressing the response value of the neighboring direction in the training correction process based on the deviation r between the neighboring direction and the labeled direction" specifically involves: establishing a functional relationship (1) between the deviation r and the second control coefficient δ2, such that the smaller the deviation r, the closer the second control coefficient δ2 is to 0 from 1.

[0017] Furthermore, in a more preferred embodiment of the present invention, the functional relationship (1) is:

[0018]

[0019] Where r∈[-R,R], representing the deviation of the current direction from the labeled direction, and σ is the standard deviation of the Gaussian distribution. To ensure that δ² can be 0, we take 0 here. e is a natural constant, and t is an adjustment coefficient for r. The closer the neighboring direction is to the labeled direction, i.e., the smaller |r| is, the closer the second adjustment coefficient δ2 is to 0 from 1, and the smaller the inhibitory effect on the response value of the neighboring direction in the training correction process.

[0020] Furthermore, in a more preferred embodiment of the present invention, the method further includes the following steps: during the model detection and inference stage, the response values ​​of K directional intervals are traversed. For each directional interval, if there is a response value in any of the D directional intervals to its left and right that is greater than the response value of the current direction, then the response value of the current directional interval is set to zero, where D is a preset parameter.

[0021] It should be noted that this operation was added because during inference and prediction, the direction of a corner point in an image may cause multiple neighboring angle intervals to have high response values. If the direction of a single corner point is directly determined by a threshold, for example, if the angle interval with a response value greater than 0.5 is considered to have a corner point direction, it will cause many false detections. This operation can retain the direction interval with the highest response value among D direction intervals and ignore the second highest direction interval. When determining the corner point direction by a threshold, such false detections can be avoided.

[0022] Furthermore, in a more preferred embodiment of the present invention, the method further includes: calculating the total loss value, updating the convolutional network model, and training is completed if the convolutional network model converges.

[0023] Furthermore, this invention also provides a method for training a parking space detection model. The method includes: acquiring sample data, the sample data including a surround view image, corner coordinates, and corner direction annotation information; inputting the surround view image into a convolutional network model; identifying normal corner positions, difficult corner positions, and non-corner positions in the surround view image based on the corner coordinate annotation information; calculating a loss value using normal corner positions as positive samples; calculating a loss value using difficult corner positions as positive samples, multiplying the loss value by a coefficient δ1∈[0,1); and calculating a loss value using non-corner positions as negative samples. The convolutional network model is divided into K directional intervals based on the corner points, and the labeled direction, neighboring direction and other directions are identified according to the corner point direction labeling information. The labeled direction is used as a positive sample to calculate the loss value. The neighboring direction is used as a negative sample to calculate the loss value, and the loss value is multiplied by the coefficient δ2∈[0,1). The other directions are used as negative samples to calculate the loss value. The total loss value is calculated and the convolutional network model is updated. The model is judged to converge. If the model converges, the training is completed. If the model does not converge, the step "inputting the toroidal image into the convolutional network model" will continue.

[0024] Furthermore, the present invention also provides an electronic device, comprising:

[0025] A computer program for executing the parking space detection model training method described above;

[0026] A memory for storing the computer program;

[0027] A processor for executing the computer program.

[0028] The parking space detection model training method provided by this invention specifically includes the following technical contents: identifying the type of parking space corner points in a parking space surround view image, wherein the parking space corner point types are divided into: normal corner points, difficult corner points, and non-corner points; if there are normal corner points in the parking space surround view image, then a normal corner point label is associated with the parking space surround view image; if there are difficult corner points in the parking space surround view image, then a difficult corner point label is associated with the parking space surround view image; in the training of the deep learning-based parking space recognition model, the parking space recognition model is trained and corrected through a loss function; if there are non-corner points in the parking space surround view image, then a non-corner point label is associated with the parking space surround view image; wherein, the normal corner point label of the parking space surround view image is used as the first positive sample, and the first positive... The loss value of the sample directly participates in the training and correction process of the parking space recognition model; the difficult corner label of the parking space surround view image is used as the second positive sample, and the loss value of the second positive sample is multiplied by the first adjustment coefficient δ1 and then participates in the training and correction process of the parking space recognition model, where the first adjustment coefficient δ1 ∈ [0,1). The non-corner label of the parking space surround view image is used as the negative sample and participates in the training and correction process of the parking space recognition model. Compared with the prior art, the technical solution of the present invention helps to reduce the negative impact of difficult targets on the model, while reducing the requirements for the judgment of the annotators and improving the clarity and feasibility of the annotation standards. The present invention also provides a parking space detection model training method and electronic device, which also have the above-mentioned beneficial effects. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0030] Figure 1 This is a flowchart illustrating the training method for a corner-based parking space detection model according to an embodiment of the present invention.

[0031] Figure 2 This is a flowchart illustrating the training method for a parking space detection model based on corner direction, as described in an embodiment of the present invention.

[0032] Figure 3 This is a flowchart illustrating the training method for the parking space detection model according to an embodiment of the present invention. Detailed Implementation

[0033] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0035] To better understand the above technical solutions, the following will describe the above technical solutions in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solutions of this application, rather than limitations on the technical solutions of this application. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0036] Specific reference Figures 1 to 3 This invention provides a parking space detection model training method, including the following technical solutions: identifying the types of parking space corner points in a parking space surround view image, wherein the parking space corner point types are divided into: normal corner points, difficult corner points, and non-corner points; if normal corner points exist in the parking space surround view image, then a normal corner point label is associated with the parking space surround view image; if difficult corner points exist in the parking space surround view image, then a difficult corner point label is associated with the parking space surround view image; in the training of the deep learning-based parking space recognition model, the parking space recognition model is trained and corrected through a loss function; if non-corner points exist in the parking space surround view image, then a non-corner point label is associated with the parking space surround view image; wherein, the normal corner point label of the parking space surround view image is used as the first positive sample. The loss value of the first positive sample directly participates in the training and correction process of the parking space recognition model; the difficult corner label of the parking space panoramic image serves as the second positive sample, and the loss value of the second positive sample, after being multiplied by the first adjustment coefficient δ1, participates in the training and correction process of the parking space recognition model, where the first adjustment coefficient δ1 ∈ [0,1). The non-corner label of the parking space panoramic image serves as the negative sample and participates in the training and correction process of the parking space recognition model. Compared with the prior art, the technical solution of the present invention helps to reduce the negative impact of difficult targets on the model, while reducing the requirements for the judgment of the annotators and improving the clarity and feasibility of the annotation standards. The present invention also provides an electronic device that has the same beneficial effects.

[0037] It should be noted that the characteristic of the difficult corner points involved in the embodiments of this invention is that "the annotator cannot determine whether this location should be labeled as a corner point." This situation may be caused by motion blur, poor lighting at night, ground reflection, partial occlusion, or image distortion and misalignment caused by perspective transformation and stitching errors. Most supervised deep learning algorithms also encounter difficult examples, but deep learning algorithms have a certain tolerance for a small number of incorrect annotations and generally do not require special handling. However, in parking space detection tasks, due to the complex working conditions, there will be a large number of ambiguous difficult examples. Implementing this invention can significantly improve the performance of parking space detection algorithms. Furthermore, the normal corner points, difficult corner points, and non-corner points mentioned in this invention are labels for the labeled targets, not labels for the entire image. An image may have multiple corner points, and their labels may be different.

[0038] Specifically, in this embodiment of the invention, the method further includes the following steps: identifying the labeled direction of the parking space corner point in the parking space surround view image; discretizing the continuous 0 to 360° into K angle intervals based on the parking space corner point, where K is a positive integer; marking the ±R angle intervals located to the left and right of the angle interval where the labeled direction is located as neighboring directions; marking other angle intervals beyond the ±R angle intervals as other directions; in the training of the deep learning-based parking space recognition model, training and correcting the parking space recognition model through a loss function; enhancing the response value of the labeled direction in the process of participating in the training and correction of the parking space recognition model; suppressing the response values ​​of the neighboring direction and the other directions in the process of participating in the training and correction of the parking space recognition model, wherein the suppression effect of the response value of the neighboring direction is proportional to the product obtained by multiplying the loss value of the neighboring direction in the training and correction process by the second control coefficient δ2, and δ2∈[0,1).

[0039] Specifically, in this embodiment of the invention, the step "suppressing the response value of the neighboring direction in the training and correction process of the parking space recognition model" specifically means: suppressing the response value of the neighboring direction in the training and correction process based on the deviation r between the neighboring direction and the labeled direction, where r∈[-R,R].

[0040] Specifically, in this embodiment of the invention, the step "suppressing the response value of the neighboring direction in the training correction process based on the deviation r between the neighboring direction and the labeled direction" specifically means: establishing a functional relationship (1) between the deviation r and the second control coefficient δ2, such that the smaller the deviation r, the closer the second control coefficient δ2 is to 0 from 1.

[0041] Specifically, in this embodiment of the invention, the functional relationship (1) is:

[0042]

[0043] Where r∈[-R,R], representing the deviation of the current direction from the labeled direction, and σ is the standard deviation of the Gaussian distribution. To ensure that δ² can be 0, we take 0 here. e is a natural constant, and t is an adjustment coefficient for r. The closer the neighboring direction is to the labeled direction, i.e., the smaller |r| is, the closer the second adjustment coefficient δ2 is to 0 from 1, and the smaller the inhibitory effect on the response value of the neighboring direction in the training correction process.

[0044] Specifically, in this embodiment of the invention, the method further includes the following steps: during the model detection and inference stage, the response values ​​of K directional intervals are traversed. For each directional interval, if there is a response value in any of the D directional intervals to its left and right that is greater than the response value of the current direction, the response value of the current directional interval is set to zero, where D is a preset parameter.

[0045] It should be noted that this operation was added because during inference and prediction, the direction of a corner point in an image may cause multiple neighboring angle intervals to have high response values. If the direction of a single corner point is directly determined by a threshold, for example, if the angle interval with a response value greater than 0.5 is considered to have a corner point direction, it will cause many false detections. This operation can retain the direction interval with the highest response value among D direction intervals and ignore the second highest direction interval. When determining the corner point direction by a threshold, such false detections can be avoided.

[0046] Specifically, in this embodiment of the invention, the method further includes: calculating the total loss value, updating the convolutional network model, and training is completed if the convolutional network model converges.

[0047] Furthermore, this invention also provides a method for training a parking space detection model. This method includes: acquiring sample data, which includes a surround view image, corner coordinates, and corner direction annotation information; inputting the surround view image into a convolutional network model; identifying normal corner positions, difficult corner positions, and non-corner positions in the surround view image based on the corner coordinate annotation information; calculating a loss value using normal corner positions as positive samples; calculating a loss value using difficult corner positions as positive samples, multiplying the loss value by a coefficient δ1∈[0,1); and calculating a loss value using non-corner positions as negative samples. Divide the image into K directional intervals based on the corner points, and identify the labeled direction, neighboring direction, and other directions based on the corner point directional labeling information. Calculate the loss value using the labeled direction as a positive sample; calculate the loss value using the neighboring direction as a negative sample, and multiply the loss value by a coefficient δ2∈[0,1); calculate the loss value using other directions as negative samples; calculate the total loss value and update the convolutional network model; determine if the model has converged. If the model has converged, training is complete; if the model has not converged, continue with the step "input the toroidal image into the convolutional network model".

[0048] In addition, the present invention also provides an electronic device, comprising: a computer program for executing the parking space detection model training method as described above; a memory for storing the computer program; and a processor for executing the computer program.

[0049] More specifically, in this embodiment of the invention, a deep learning-based method for detecting parking space corners is proposed. The specific corner processing steps are as follows:

[0050] 1. During the manual annotation stage, ambiguous corner points are marked as difficult examples;

[0051] 2. During the training phase, when calculating the loss function, hard examples are also treated as positive samples, but the loss value is multiplied by a coefficient to reduce the impact of hard examples as positive samples on the model. One feasible solution is to set δ1 = 0 to exclude the impact of hard examples as positive samples on the model.

[0052] 3. During the detection phase, normal detection and reasoning are sufficient.

[0053] The advantages of the above processing are as follows: The effective features of hard examples, as positive samples, are not suppressed; and the impact of their pseudo-features on the model is limited when their loss value is multiplied by a coefficient less than 1. Specifically, when δ1 = 0, neither the effective features nor the pseudo-features of the hard examples affect the model. Furthermore, the above processing method improves the executability and efficiency of annotation standards. In complex situations, annotation standards often require numerous detailed implementation rules, which can affect executability, annotation efficiency, and accuracy. Introducing hard examples is equivalent to setting up a buffer between positive and negative samples. For ambiguous situations, the cost of confusing positive samples with hard examples or negative samples with hard examples is very low and will not affect model convergence.

[0054] The steps for handling the direction of corner points are as follows:

[0055] 1. During the manual labeling stage, for curved, blurry, or misaligned parking lines, simply label their general direction.

[0056] 2. During the training phase, increase the response value in the labeled direction and suppress the response value in other directions. For each of the R directions to the left and right of the labeled direction, the loss value should be multiplied by a coefficient to reduce the suppression of directions adjacent to the labeled direction, where R is a preset parameter. One possible implementation is to take...

[0057]

[0058] Where r∈[-R,R], representing the deviation of the current direction from the labeled direction, and σ is the standard deviation of the Gaussian distribution. To ensure that δ² can be 0, we take 0 here. e is a natural constant, and t is an adjustment coefficient for r. The closer the neighboring direction is to the labeled direction, i.e., the smaller |r| is, the closer the second adjustment coefficient δ2 is to 0 from 1, and the smaller the inhibitory effect on the response value of the neighboring direction in the training correction process.

[0059] 3. During the detection phase, the training method in step 2 may result in high responses in multiple consecutive directional intervals. Non-maximum suppression is required to select the direction with the highest response value.

[0060] Specifically, for each direction, if there is a larger response in each of the D directions to its left and right, then the response value for that direction is set to 0, where D is a preset parameter.

[0061] The advantage of the above treatment is that it reduces the impact of labeling bias on model convergence. The training objective is for labeled directions to have higher response values ​​and unlabeled directions to have lower response values. Labeling bias often causes the training objective in some directions to be reversed, thus affecting model convergence. In this invention, the loss value of the directions adjacent to the labeled direction is multiplied by a coefficient δ2, reducing the response values ​​in these directions. Furthermore, the coefficient δ2 gradually approaches 0 the closer to the labeled direction, and the suppression effect on the response value also gradually approaches 0. Specifically, when calculating the loss function, it is necessary to distinguish between positive and negative samples. The loss function objective for positive samples is to increase the response value, while the loss function objective for negative samples is to suppress the response value. The adjacent directions are considered negative samples, and their response values ​​are to be suppressed. The smaller the coefficient, the smaller the suppression effect, allowing high response values ​​in the labeled direction and its adjacent directions. Therefore, the impact of labeling bias on model convergence is significantly reduced.

[0062] Furthermore, the above processing reduces the difficulty of annotation and improves annotation efficiency. For curved, blurred, or misaligned parking lines, establishing annotation standards for corner directions is often very tedious and difficult to implement accurately, yet it is crucial because unclear annotation standards or errors in annotation personnel's judgment can lead to contradictory annotation results, affecting model convergence. The corner direction processing method in this invention can tolerate minor annotation deviations in similar scenarios, and the annotation standards do not need to be cumbersomely specified for various complex working conditions.

[0063] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0064] Similarly, it should be understood that, in order to simplify this disclosure and aid in understanding one or more of the various aspects of the invention, in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof. However, this method of disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into this detailed description, wherein each claim itself is a separate embodiment of the invention.

[0065] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature serving the same, equivalent, or similar purpose.

[0066] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0067] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

Claims

1. A method for training a parking space detection model, characterized in that, Includes the following steps: Identify the type of parking space corner points in the parking space surround view image. The types of parking space corner points are divided into: normal corner points, difficult corner points, and non-corner points. If there are normal corner points in the parking space surround view image, then associate normal corner point labels with the parking space surround view image; and If there are difficult corner points in the parking space surround view image, then the difficult corner point label is associated with the parking space surround view image; and If there are non-corner points in the parking space surround view image, then the non-corner point label is associated with the parking space surround view; In the training of the deep learning-based parking space recognition model, the model is trained and corrected using a loss function. The normal corner labels of the parking space surround view image are used as the first positive samples, and the loss value of the first positive samples directly participates in the training and correction process of the parking space recognition model. The difficult corner label of the parking space surround view image is used as the second positive sample. The loss value of the second positive sample is multiplied by the first adjustment coefficient δ1 and then participates in the training and correction process of the parking space recognition model. The first adjustment coefficient δ1∈[0,1). The non-corner labels of the parking space surround view image are used as negative samples in the training and correction process of the parking space recognition model. The annotation direction of the parking space corner point in the parking space surround view image is identified. Based on the parking space corner point, the continuous 0 to 360° is discretized into K angle intervals, where K is a positive integer. The ±R angle intervals that are adjacent to the left and right of the angle interval where the annotation direction is located are marked as adjacent directions. Other angle intervals that are beyond the ±R angle intervals are marked as other directions. In the training of the deep learning-based parking space recognition model, the parking space recognition model is trained and corrected through a loss function. The response value of the annotation direction in the process of participating in the training and correction of the parking space recognition model is enhanced. The response values ​​of the adjacent direction and the other directions in the process of participating in the training and correction of the parking space recognition model are suppressed. The suppression effect of the response value of the adjacent direction is proportional to the product of the loss value of the adjacent direction in the training and correction process and the second control coefficient δ2, and δ2∈[0,1). The step "suppress the response value of the neighboring direction in the training and correction process of the parking space recognition model" specifically means: suppressing the response value of the neighboring direction in the training and correction process based on the deviation r between the neighboring direction and the labeled direction, where r∈[-R,R].

2. The parking space detection model training method according to claim 1, characterized in that, The step "suppressing the response value of the neighboring direction in the training correction process based on the deviation r between the neighboring direction and the labeled direction" specifically involves: establishing a functional relationship (1) between the deviation r and the second control coefficient δ2, such that the smaller the deviation r, the closer the second control coefficient δ2 is to 0 from 1.

3. The parking space detection model training method according to claim 2, characterized in that, The functional relationship (1) is as follows: Where r∈[-R,R], representing the deviation of the current direction from the labeled direction, and σ is the standard deviation of the Gaussian distribution. To ensure that δ² can be 0, we take 0 here. e is a natural constant, and t is an adjustment coefficient for r. The closer the neighboring direction is to the labeled direction, i.e., the smaller |r| is, the closer the second adjustment coefficient δ2 is to 0 from 1, and the smaller the inhibitory effect on the response value of the neighboring direction in the training correction process.

4. The parking space detection model training method according to claim 3, characterized in that, It also includes the following steps: In the model detection and inference stage, the response values ​​of K directional intervals are traversed. For each directional interval, if there is a response value in the D directional intervals to its left and right that is greater than the response value of the current direction, then the response value of the current directional interval is set to zero, where D is a preset parameter.

5. The parking space detection model training method according to any one of claims 1 to 4, characterized in that, Also includes: Calculate the total loss value, update the convolutional network model, and the training is complete if the convolutional network model converges.

6. A method for training a parking space detection model, characterized in that, The method includes: Obtain sample data, which includes a panoramic image, corner coordinates, and corner direction annotations. The toroidal image is input into the convolutional network model; Based on the corner coordinate annotation information, the normal corner position, difficult corner position, and non-corner position in the panoramic image are identified; the normal corner position is used as a positive sample to calculate the loss value; the difficult corner position is used as a positive sample to calculate the loss value, and the loss value is multiplied by the coefficient δ1∈[0,1); the non-corner position is used as a negative sample to calculate the loss value. Divide the area into K directional intervals based on the corner points, and identify the labeled direction, neighboring direction and other directions based on the corner point directional labeling information; calculate the loss value using the labeled direction as a positive sample; calculate the loss value using the neighboring direction as a negative sample, and multiply the loss value by a coefficient δ2∈[0,1); calculate the loss value using other directions as negative samples. Calculate the total loss value mentioned above and update the convolutional network model; Determine if the model has converged. If the model has converged, the training is complete. If the model has not converged, continue to the step "input the toroidal image into the convolutional network model".

7. An electronic device, characterized in that, include: A computer program, the computer program being used to execute the parking space detection model training method according to any one of claims 1-5; A memory for storing the computer program; A processor for executing the computer program.

Images