Visual parking space detection method, system and electronic device based on direction angle point

Abstract

This invention discloses a visual parking space detection method based on directional corner points, comprising: inputting a surround view image of the vehicle to be detected into a model; obtaining parking space status data and parking space corner point data from the model; obtaining parking space information based on the parking space status data and parking space corner point data; the parking space status data includes: parking space boundary position data and parking space attribute data; the parking space corner point data includes: parking space corner point coordinate data and adjacent corner point direction data; wherein, the model is obtained by deep learning training using multiple sets of data. The technical solution of this invention uses directional corner points as the basic unit of parking spaces. Compared with the currently commonly used methods that use corner points or parking lines as basic units, its advantages are that the features of corner points and their directions are more explicit, the versatility is stronger, and parking space detection can be completed accurately and efficiently. This invention also provides a visual parking space detection system and an electronic device based on directional corner points, which also have the above-mentioned beneficial effects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, in particular to a visual parking space detection method based on a direction angle point, and the present application also relates to a visual parking space detection system based on a direction angle point and an electronic device, which also have the above beneficial effects. BACKGROUND

[0002] With the development of technology and the continuous improvement of regulations, automatic driving technology is gradually landing, and automatic parking is one of the current research focuses of various vehicle manufacturers. Parking space detection is one of the key perception tasks in the automatic parking system. The basic implementation method is that the parking space detection module detects the specific position of the parking space according to the surrounding environment information collected by the vehicle camera or ultrasonic radar sensor, and the subsequent module controls the vehicle to park according to the parking position.

[0003] The key of the parking space detection technology is accurate positioning and strong universality. However, in the existing visual parking space detection method, the edge information of the image is used to extract the parking line, which leads to a serious performance decline for parking spaces with shadows or unevenness. In addition, the existing technology also uses a semantic segmentation method to extract the parking line, and then combines the extraction of the parking space according to the semantic segmentation result. The image distortion caused by factors such as the width of the parking line, other marking lines, fisheye cameras or stitching errors in the actual scene will affect the success rate of parking space positioning, and the workload of training data labeling and segmentation result post-processing is very large. SUMMARY

[0004] To solve the above technical problems, the main purpose of the present application is to provide a visual parking space detection method based on a direction angle point, which can accurately and efficiently complete parking space detection. The present application also provides a visual parking space detection system based on a direction angle point and an electronic device, which also have the above beneficial effects.

[0005] To achieve the above purpose, the present application provides a visual parking space detection method based on a direction angle point, which includes the following contents:

[0006] Input the to-be-detected vehicle body panoramic image into the model;

[0007] Obtain parking space state data and parking angle point data from the model;

[0008] Obtain parking space information based on the parking space state data and the parking angle point data;

[0009] The parking space state data includes parking boundary position data and parking attribute data, and the parking angle point data includes parking angle point coordinate data and adjacent angle point direction data.

[0010] The model is trained by deep learning using multiple sets of data, each set of training data including: a training vehicle surround view image, a first label indicating the position of the parking space boundary in the image, a second label indicating the attribute of the parking space in the image, a third label indicating the coordinates of the corner points of the parking space in the image, and a fourth label indicating the directions of all adjacent corner points of the corner points of the parking space in the image.

[0011] Further, in the visual parking space detection method based on the direction angle point provided by the application, the model is trained by a convolution network; the convolution network includes: a first output head for outputting parking space state data; and a second output head for outputting parking space corner point data.

[0012] Further, in the visual parking space detection method based on the direction angle point provided by the application, the parking space attribute includes: non-parking space, empty parking space, occupied parking space, and / or the parking space attribute includes: parking space with activated ground lock, empty parking space with conical barrel.

[0013] Further, in the visual parking space detection method based on the direction angle point provided by the application, the parking space attribute includes: parking space with activated ground lock, empty parking space with conical barrel.

[0014] Further, in the visual parking space detection method based on the direction angle point provided by the application, the first output head includes a parking space attribute type output channel for outputting the parking space attribute data, and a parking space boundary position output channel for outputting the parking space boundary position data; the second output head includes an adjacent corner point direction output channel for outputting the adjacent corner point direction data, and a parking space corner point coordinate output channel for outputting the parking space corner point coordinate data; wherein the number of parking space attribute type output channels is N channels, N is the number of parking space attributes; the number of parking space boundary position output channels is 4 channels; the number of adjacent corner point direction output channels is K channels, K is the number of angle segments into which 360° is divided, and each channel represents an angle interval; and the number of parking space corner point coordinate output channels is 1 channel.

[0015] Further, in the visual parking space detection method based on the direction angle point provided by the application, it further includes: calling the front, rear, left and right cameras on the vehicle to take pictures of the environment to obtain a to-be-detected vehicle surround view image, including: acquiring a first original image collected by the camera; and splicing multiple first original images to obtain the to-be-detected vehicle surround view image.

[0016] Further, in the visual parking space detection method based on the direction angle point provided by the application, the method for making the training data comprises: acquiring the training vehicle body surround view image; analyzing the training vehicle body surround view image; identifying the parking space boundary to generate the first label of the parking space boundary position in the training vehicle body surround view image; identifying the parking space attribute to generate the second label of the parking space attribute in the training vehicle body surround view image; identifying the parking space angle point to generate the third label of the parking space angle point in the training vehicle body surround view image; and identifying the adjacent angle point direction to generate the fourth label of the adjacent angle point direction in the training vehicle body surround view image.

[0017] Further, in the visual parking space detection method based on the direction angle point provided by the application, the method further comprises: dynamically adjusting all weights of the model by using a loss function; and the loss function is:

[0018] L = L cls + αL reg + βL p + γL dir (1).

[0019] Wherein, L represents the loss function, L cls represents the parking space attribute classification loss; L reg represents the regression loss of the parking space boundary position; L p represents the loss of the parking space angle point coordinate prediction; L dir represents the adjacent angle point direction prediction loss; and α, β, γ represent the coefficients of different parts of the loss.

[0020] In addition, the application also provides a visual parking space detection system based on the direction angle point, which comprises: a first transmission module for inputting a vehicle body surround view image to be detected into a model; a second acquisition module for acquiring parking space state data and parking space angle point data from the model; and a third analysis module for obtaining parking space information based on the parking space state data and the parking space angle point data.

[0021] Wherein, the model is obtained by using multiple sets of data through deep learning training, and each set of training data in the multiple sets of data comprises: a training vehicle body surround view image, a first label indicating the parking space boundary position in the image, a second label indicating the parking space attribute in the image, a third label indicating the parking space angle point coordinate in the image, and a fourth label indicating the adjacent angle point direction of the parking space angle point in the image.

[0022] The parking space state data comprises: parking space boundary position data and parking space attribute data.

[0023] The parking space angle point data comprises: parking space angle point coordinate data and adjacent angle point direction data.

[0024] The parking space attribute includes: non-parking space, empty parking space, occupied parking space; and / or

[0025] The parking space attribute includes: parking space with a starting ground lock, empty parking space with a conical barrel.

[0026] Further, in the visual parking space detection system based on the direction angle point provided by the application, the system further comprises a fourth acquisition module for calling a camera on a vehicle to take a picture of the environment to obtain a to-be-detected vehicle surround view image.

[0027] In addition, the application also provides an electronic device, comprising: software for executing the visual parking space detection method based on the direction angle point; a memory for storing the software; and a processor for executing the software.

[0028] The visual parking space detection method based on the direction angle point provided by the application specifically comprises the following technical contents: inputting a to-be-detected vehicle surround view image into a model; obtaining parking space state data and parking space angle point data from the model; obtaining parking space information based on the parking space state data and the parking space angle point data; the parking space state data includes parking space boundary position data and parking space attribute data, and the parking space angle point data includes parking space angle point coordinate data and adjacent angle point direction data; wherein the model is obtained by using multiple sets of data for deep learning training, each set of training data in the multiple sets of data includes a training vehicle surround view image, a first label indicating the parking space boundary position in the image, a second label indicating the parking space attribute in the image, a third label indicating the parking space angle point coordinate in the image, and a fourth label indicating the direction of all adjacent angle points of the parking space angle point; the parking space attribute includes non-parking space, empty parking space, and occupied parking space. In the technical solution involved in the application, the angle point with direction is used as the basic unit of the parking space, compared with the current commonly used angle point or parking space line as the basic unit method, the features of the angle point and its direction are more explicit, and the universality is stronger. Compared with the prior art, the technical solution involved in the application can accurately and efficiently complete the parking space detection. In addition, the application also provides a visual parking space detection system based on the direction angle point and an electronic device, which also have the above beneficial effects. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only embodiments of the application, and those skilled in the art can obtain other drawings according to the provided drawings without any creative effort:

[0030] Figure 1A summary flow chart of a visual parking space detection method based on a belt direction angle point according to the present application;

[0031] Figure 2 A visual parking space detection model schematic diagram according to the present application;

[0032] Figure 3 A model training flow chart schematic diagram according to the present application;

[0033] Figure 4 A corner point direction schematic diagram according to the present application. DETAILED DESCRIPTION

[0034] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The drawings show typical embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided for the purpose of making the disclosure of the present application more thorough and comprehensive.

[0035] 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 the present application belongs. The terms used herein in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0036] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings and specific embodiments of the specification. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solutions of the present application, and are not intended to limit the technical solutions of the present application. In the case of no conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0037] Specifically refer to Figures 1 to 4The application provides a visual parking space detection method based on a direction angle point, and specifically comprises the following technical contents: inputting a to-be-detected vehicle body surround view image into a model; obtaining parking space state data and parking space angle point data from the model; obtaining parking space information based on the parking space state data and the parking space angle point data; the parking space state data comprises parking space boundary position data and parking space attribute data, and the parking space angle point data comprises parking space angle point coordinate data and adjacent angle point direction data; wherein the model is obtained by using multiple sets of data to train through deep learning, each set of training data in the multiple sets of data comprises a training vehicle body surround view image, a first label indicating a parking space boundary position in the image, a second label indicating a parking space attribute in the image, a third label indicating a parking space angle point coordinate in the image, and a fourth label indicating all adjacent angle point directions of the parking space angle point in the image; the technical scheme disclosed in the application takes a direction angle point as a basic unit of a parking space, and compared with a commonly used angle point or parking space line as a basic unit method, the angle point and its direction feature are more explicit and more universal; compared with the prior art, the technical scheme disclosed in the application can accurately and efficiently complete parking space detection; in addition, the application also provides a visual parking space detection system based on a direction angle point and an electronic device, which also have the above beneficial effects.

[0038] Specifically, in the embodiment of the application, the model is trained by a convolution network; the convolution network comprises: a first output head for outputting parking space state data; and a second output head for outputting parking space angle point data.

[0039] Specifically, in the embodiment of the application, the parking space attribute comprises: a non-parking space, an empty parking space, and an occupied parking space.

[0040] Specifically, in the embodiment of the application, the parking space attribute comprises: a parking space with a starting ground lock and an empty parking space with a conical barrel.

[0041] Specifically, in the embodiment of the present application, the first output head comprises a parking space attribute type output channel for outputting the parking space attribute data and a parking space boundary position output channel for outputting the parking space boundary position data; the second output head comprises an adjacent corner point direction output channel for outputting the adjacent corner point direction data and a parking space corner point coordinate output channel for outputing parking space corner point coordinate data; the data in the channels and the input image each pixel position are one-to-one corresponding, and each data in the channels represents the probability that the corresponding position pixel is a corner point. Wherein, the number of the parking space attribute type output channel is N channels, N is the number of parking space attributes; the number of the parking space boundary position output channel is 4 channels; the number of the adjacent corner point direction output channel is K channels, K is the number of angle parts for dividing 360°, and each channel refers to an angle interval; the number of the parking space corner point coordinate output channel is 1 channel.

[0042] Specifically, in the embodiment of the present application, further comprising: calling the cameras in the front, rear, left and right directions of the vehicle to take pictures of the environment to obtain the to-be-detected vehicle surround view image, comprising: acquiring a first original image collected by the camera; splicing a plurality of the first original images to obtain the to-be-detected vehicle surround view image.

[0043] Specifically, in the embodiment of the present application, the method for making the training data comprises: acquiring the training vehicle surround view image; analyzing the training vehicle surround view image; identifying the parking space boundary to generate a first label of the parking space boundary position in the training vehicle surround view image; identifying the parking space attribute to generate a second label of the parking space attribute in the training vehicle surround view image; identifying the parking space corner point to generate a third label of the parking space corner point in the training vehicle surround view image; and identifying the adjacent corner point direction to generate a fourth label of the adjacent corner point direction in the training vehicle surround view image.

[0044] Specifically, in the embodiment of the present application, the method further comprises: dynamically adjusting all weights of the model by using a loss function; and the loss function is:

[0045] L=L cls +αL reg +βL p +γL dir (1);

[0046] Wherein, L represents a loss function, L cls represents a parking space attribute classification loss; L reg represents a parking space boundary position regression loss; L p represents a parking space corner point coordinate prediction loss; and L dirindicates the adjacent corner point direction prediction loss; and a, b, and g indicate different coefficients of the loss of each part.

[0047] In addition, the application also provides a visual parking space detection system based on a corner point with a direction, comprising: a first transmission module for inputting a to-be-detected vehicle surround view image into a model; a second acquisition module for acquiring parking space state data and parking space corner point data from the model; and a third analysis module for obtaining parking space information based on the parking space state data and the parking space corner point data.

[0048] The model is obtained by deep learning training using a plurality of sets of data, and each set of training data in the plurality of sets of data comprises: a training vehicle surround view image, a first label indicating a parking space boundary position in the image, a second label indicating a parking space attribute in the image, a third label indicating a parking space corner point coordinate in the image, and a fourth label indicating all adjacent corner point directions of the parking space corner point in the image.

[0049] The parking space state data comprises: parking space boundary position data and parking space attribute data.

[0050] The parking space corner point data comprises: parking space corner point coordinate data and adjacent corner point direction data.

[0051] The parking space attribute comprises: a non-parking space, an empty parking space, and an occupied parking space; and / or the parking space attribute comprises: a parking space with a starting ground lock and an empty parking space with a conical barrel.

[0052] Specifically, in the embodiment of the application, the system further comprises a fourth acquisition module for calling a camera on a vehicle to take a picture of the environment to obtain the to-be-detected vehicle surround view image.

[0053] In addition, the application also provides an electronic device, comprising: software for executing the visual parking space detection method based on a corner point with a direction as described above; a memory for storing the software; and a processor for executing the software.

[0054] More specifically from the overall inventive concept, the technical solution related to the embodiment of the application belongs to the field of intelligent driving, and specifically relates to a visual parking space detection method based on a corner point with a direction, which can be used for detecting and positioning a parking space by an automatic parking system.

[0055] With the development of technology and the continuous improvement of regulations, automatic driving technology is gradually landing, and automatic parking is one of the current research focuses of various vehicle manufacturers. Parking space detection is one of the key perception tasks in the automatic parking system, and the basic implementation mode is that a parking space detection module detects the specific position of a parking space according to the surrounding environment information collected by a vehicle camera or an ultrasonic radar sensor, and a subsequent module controls the vehicle to park into the parking space according to the parking space position.

[0056] The key of parking space detection technology is positioning accuracy and strong universality, however, in the prior art, the visual parking space detection method is mostly to extract the parking line by using image edge information, and the performance is seriously reduced for the parking space with light shadow or unevenness; in addition, the prior art also extracts the parking line by using a semantic segmentation method, and then combines the parking space according to the semantic segmentation result, and the image distortion caused by the width of the parking line, other identification lines, a fisheye camera or splicing error and other factors in the actual scene will affect the success rate of the parking space positioning, and the workload of the annotation of the training data and the post-processing of the segmentation result is large.

[0057] Accordingly, the present application provides a visual parking space detection method based on a direction angle point, mainly comprising the following steps: pre-using a manually labeled sample picture to supervise the training of a convolution network to obtain a visual parking space detection model; in a detection stage, inputting a to-be-detected vehicle panoramic image into the model, and then analyzing parking space coordinates according to an output result, wherein the convolution network has two output heads, a first output head is a standard target detection task detection head, and is used to output a parking space bounding box and a classification label, specifically, the label can be divided into three categories, namely: non-parking space, empty parking space and occupied parking space.

[0058] Specifically, the first output head has N+4 channels, wherein N is the number of label categories, indicating the probability of belonging to each category, and the last four channels are position regression maps, indicating the four boundary positions of the parking space bounding box.

[0059] The second output head is used to output the angle point coordinates of the parking space and the direction of all adjacent angle points.

[0060] Specifically, in the technical solution, the direction of the angle point is represented in the form of quantitative classification, and 360 degrees is equally divided into K intervals, and each interval is a category, the second output head has 1+K channels, wherein the first channel represents the probability of the angle point, and the last K channels are each taken as a binary classification problem, and represent the probability of the existence of the corresponding direction.

[0061] More specifically, in the embodiment of the present application, K is preferably 90, and each angle span is 4 degrees, and the error is within 2 degrees.

[0062] It should be noted that taking the corner point with direction as the basic unit of the parking space is an important point of the present application, and compared with the method of taking the corner point or the parking space line as the basic unit, the corner point with direction has more clear features and higher universality. In actual working conditions, there are non-parking space lines, and some parking space lines are not complete. The input for parking space detection is usually a surround view formed by switching the view angle of four fisheye cameras, and parameter errors may cause the actual straight line to be curved, blurred, misaligned, etc. In these cases, the corner point direction is more clear than the straight line, and is more conducive to the convergence and prediction of the model.

[0063] The corner point with direction as the basic unit of the parking space is simple and efficient for constructing the parking space. In the prior art, the parking space line usually needs to be extracted by edge detection or extracted from the semantic segmentation result, which has a large amount of calculation. In comparison, the corner point with direction can be matched to construct the parking space, which is simple and efficient.

[0064] In addition, designing the detection head of the parking space bounding box and the detection head of the corner point with direction in one convolutional network is also an important point of the present application, and has the following advantages:

[0065] First, the parking space detection head determines the parking space region and the label of whether it can be parked, and the corner point with direction locates the coordinates of the parking space in the region. The two detection heads are organically combined to efficiently complete the parking space detection task.

[0066] Second, the features of the corner point with direction and the features of the parking space are highly coupled, the two detection heads are integrated into one convolutional network, and the two kinds of supervision information complement each other, so that the model can understand the task more comprehensively, achieving the effect of 1 plus 1 greater than 2.

[0067] The visual parking space detection model is obtained by supervised training, and the supervision information comes from a large number of manually labeled sample pictures. According to the introduction of the convolutional network, the labeling information needs to include the parking space bounding box, the corner point coordinates and the direction.

[0068] The corner point direction is represented by the polar angle of the adjacent corner point, and is specifically shown in Figure 4 If the adjacent corner point in a certain direction is a visual breakpoint, any point in the corresponding direction is used instead.

[0069] In actual labeling, the virtual point and the corner point do not need to be distinguished. The virtual point has only one direction, and the corner point has two or more directions. In the training stage, the virtual point only provides the corner point direction calculation for its adjacent corner points, and does not supervise the model in other ways. In order to reduce the direction labeling error, the virtual point should not be selected too close to the corresponding corner point.

[0070] In the model training stage, the two output heads of the network are supervised trained by using the label information. It is particularly pointed out that the direction value of the corner point is calculated according to all the adjacency relations of the corner point, and then the direction value is mapped into the corresponding category in K categories, so that it can be directly used for supervision of the angle direction prediction. The loss function is:

[0071] L = L cls + αL reg + βL p + γL dir (1)

[0072] Wherein L represents the loss function, represents the parking space frame classification loss, including but not limited to using cross entropy loss; represents the regression loss of the four edges of the parking space frame, including but not limited to using L1 loss, L2 loss, IOU loss, DIOU loss; represents the loss of the corner point prediction, including but not limited to using cross entropy loss; represents the direction prediction loss, including but not limited to using cross entropy loss; α, β, γ represent the coefficients of different parts of the loss.

[0073] In the specification provided herein, a large number of specific details are described. However, it can be understood that the embodiments of the application can be practiced without these specific details. In some examples, well-known methods, structures and techniques are not shown in detail in order not to obscure the understanding of the present specification.

[0074] Similarly, it should be appreciated that the individual features of the application have sometimes been presented together in a single embodiment, drawing or description of an embodiment in the foregoing description of the application's exemplary embodiments to simplify the disclosure and to help understand one or more of the various inventive aspects. However, the method of the disclosure should not be interpreted as reflecting an intention that the claimed application requires more features than those explicitly recited in each claim. Rather, as reflected in the following claims, the inventive aspects are in less than all the features of the single embodiment disclosed earlier. Therefore, the claims following the specific embodiments are hereby expressly incorporated into this specific embodiment, wherein each claim itself is a separate embodiment of the application.

[0075] Those skilled in the art will appreciate that the modules in the apparatuses in the embodiments can be adapted and placed in one or more apparatuses other than the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and further can be split into multiple sub-modules or sub-units or sub-components. Any combination of all the features disclosed in the specification (including the accompanying claims, abstract and drawings), and any method or process or device of any combination of the features disclosed in the specification (including the accompanying claims, abstract and drawings) can be taken, except that at least some of such features and / or processes or units are mutually exclusive, unless specifically stated otherwise. Each feature disclosed in the specification (including the accompanying claims, abstract and drawings) can be replaced by alternative features serving the same, equivalent or similar purpose, unless specifically stated otherwise.

[0076] Furthermore, those skilled in the art will appreciate that the features of the different embodiments can be combined in any combination, which is within the scope of the present application and forms a different embodiment. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0077] It is noted that the above-mentioned embodiments illustrate rather than limit the application, and that one skilled in the art will be able to design many alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word 'comprising' does not exclude the presence of elements or steps other than those listed in a claim. The word 'a' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of both hardware and software, and any combination thereof. In a unit claim, any reference signs placed between parentheses in the claim do not limit the claim. The use of the word 'at least' followed by a list of one or more items does not exclude additional such items. The use of the words 'one' or 'the' to refer to an element or an action does not exclude the presence of a plurality of such elements or actions. The use of ordinal numbers, such as 'first','second', 'third' etc., does not limit the scope of the application, but is used for the sake of nomenclature.

Claims

1. A visual parking space detection method based on band direction angle point, characterized in that, The method comprises the following steps: inputting a to-be-detected vehicle body surround view image into a model; obtaining parking space state data and parking space corner point data from the model; obtaining parking space information based on the parking space state data and the parking space corner point data; the parking space state data comprises parking space boundary position data and parking space attribute data; the parking space corner point data comprises parking space corner point coordinate data and adjacent corner point direction data; wherein the model is obtained by deep learning training using multiple sets of data, each set of training data in the multiple sets of data comprising a training vehicle body surround view image, a first label indicating a parking space boundary position in the image, a second label indicating a parking space attribute in the image, a third label indicating a parking space corner point coordinate in the image, and a fourth label indicating all adjacent corner point directions of the parking space corner point in the image; the model is trained by a convolution network, and the convolution network comprises: a first output head for outputting parking space state data; a second output head for outputting parking space corner point data; the parking space attribute comprises non-parking space, empty parking space and occupied parking space; and / or the parking space attribute comprises a parking space with a ground lock and an empty parking space with a conical barrel; the first output head comprises a parking space attribute type output channel and a parking space boundary position output channel; the parking space attribute type output channel is used for outputting the parking space attribute data, and the parking space boundary position output channel is used for outputting the parking space boundary position data; the second output head comprises an adjacent corner point direction output channel and a parking space corner point coordinate output channel, the adjacent corner point direction output channel is used for outputting the adjacent corner point direction data, and the parking space corner point coordinate output channel is used for outputting the parking space corner point coordinate data; wherein the number of the parking space attribute type output channels is N, and N is the number of parking space attributes; the number of the parking space boundary position output channels is 4; the number of the adjacent corner point direction output channels is K, K is the number of angle parts obtained by dividing 360°, and each channel represents an angle interval; the number of the parking space corner point coordinate output channels is 1.

2. The lane-direction angle point based visual parking space detection method according to claim 1, wherein, The method further comprises: calling cameras in front, back, left and right directions on a vehicle to take pictures of the environment to obtain the to-be-detected vehicle body surround view image, which comprises: obtaining a first original image collected by the cameras; splicing multiple first original images to obtain the to-be-detected vehicle body surround view image; and / or the training vehicle body surround view image is obtained by pre-taking pictures of the environment by the cameras in front, back, left and right directions on the vehicle.

3. The band direction angle point based visual parking space detection method according to claim 1, wherein, The method for making the training data comprises: obtaining the training vehicle body surround view image; analyzing the training vehicle body surround view image; identifying a parking space boundary to generate a first label of a parking space boundary position in the training vehicle body surround view image; identifying a parking space attribute to generate a second label of a parking space attribute in the training vehicle body surround view image; identifying a parking space corner point to generate a third label of a parking space corner point in the training vehicle body surround view image; and identifying an adjacent corner point direction to generate a fourth label of an adjacent corner point direction in the training vehicle body surround view image.

4. The band direction angle point based visual parking space detection method according to claim 1, wherein, The method further comprises dynamically adjusting all weights of the model by using a loss function, wherein the loss function is: (1)； wherein L denotes a loss function, denotes a parking space attribute classification loss; denotes a regression loss of a parking space boundary position; denotes a loss of a parking space corner point coordinate prediction; denotes a loss of an adjacent corner point direction prediction; , , denotes a coefficient of different partial losses.

5. A vision-based parking space detection system based on band direction angle points, characterized by, The system is used to execute the visual parking space detection method based on the direction angle point according to any one of claims 1 to 4; the system comprises: a first transmission module for inputting a to-be-detected vehicle surround view image into a model; a second acquisition module for acquiring parking space state data and parking space angle point data from the model; and a third analysis module for obtaining parking space information based on the parking space state data and the parking space angle point data. The model is obtained by deep learning training using a plurality of sets of data, and each set of training data in the plurality of sets of data comprises: a training vehicle surround view image, a first label indicating a parking space boundary position in the image, a second label indicating a parking space attribute in the image, a third label indicating a parking space angle point coordinate in the image, and a fourth label indicating a direction of all adjacent angle points of the parking space angle point in the image. The parking space state data comprises: parking space boundary position data and parking space attribute data. The parking space angle point data comprises: parking space angle point coordinate data and adjacent angle point direction data. The parking space attribute comprises: a non-parking space, an empty parking space, and an occupied parking space; and / or The parking space attribute comprises: a parking space with a starting ground lock and an empty parking space with a conical barrel. The system further comprises a fourth acquisition module for calling a camera on a vehicle to take a picture of an environment to obtain a to-be-detected vehicle surround view image.

6. An electronic device, comprising: The software is used to execute the visual parking space detection method based on the direction angle point according to any one of claims 1 to 4; The memory is used to store the software; The processor is used to execute the software. ​

Images