Parking space detection method and parking space detection system

Abstract

The application relates to a parking space detection method and system. The method comprises the following steps: collecting parking lot images containing one or more idle and occupied parking spaces around a vehicle; inputting the parking lot images into a neural network model for image processing to obtain center point coordinates and a center point probability of a parking space, first coordinates and a first corner point probability of four corner points, and offset amounts of the four corner points relative to the center point coordinates; calculating second coordinates of the four corner points according to the center point coordinates and the offset amounts of the four corner points relative to the center point coordinates; correcting the second coordinates of the four corner points by using the first coordinates of the four corner points to obtain actual coordinates and an actual corner point probability of the four corner points; and performing parking space detection based on the center point coordinates, the center point probability, and the actual coordinates and the actual corner point probability of the four corner points. According to the application, parking spaces can be accurately detected.

Description

TECHNICAL FIELD

[0001] The present application relates to automatic driving technology, in particular to a parking space detection method based on parking space corner points and a parking space detection system. BACKGROUND

[0002] The automatic driving technology, also known as the unmanned driving technology, is an intelligent automobile technology for realizing unmanned driving through a computer system. The automatic driving technology relies on the cooperation of artificial intelligence, visual computing, radar, monitoring devices and global positioning system, so that the computer system can automatically and safely operate the vehicle without any human initiative.

[0003] The automatic parking technology refers to that the vehicle is automatically parked into a parking space without manual control. Different automatic parking systems adopt different methods to detect the objects around the vehicle. Some install sensors in front and back of the vehicle, and some install cameras or radars on the vehicle to detect obstacles. The vehicle can detect the parked vehicle, the size of the parking space and the distance from the curb, and then park the vehicle into the parking space.

[0004] At present, the radar-based automatic parking system can detect the parking space between two vehicles, but cannot handle the painted parking space on the empty land. In addition, the radar may misdetect the limiters in the parking space, resulting in that the vehicle cannot completely enter the parking space and the parking is terminated. Compared with the radar-based automatic parking system, the visual-based parking space detection algorithm supports detecting the painted parking space on the empty land. SUMMARY

[0005] In view of the above problems, the present application aims to provide a parking space detection method and a parking space detection system which can accurately detect parking spaces.

[0006] The parking space detection method of one aspect of the present application is characterized by comprising:

[0007] An image acquisition step of acquiring a parking lot image containing one or more idle and occupied parking spaces around the vehicle;

[0008] An image processing step of inputting the parking lot image into a neural network model for image processing to obtain a center point coordinate and a center point probability of the parking space, a first coordinate and a first corner point probability of four corner points, and an offset of the four corner points relative to the center point coordinate;

[0009] A coordinate calculation step of calculating a second coordinate of the four corner points according to the center point coordinate and the offset of the four corner points relative to the center point coordinate;

[0010] A coordinate correction step of correcting the second coordinate of the four corner points by using the first coordinate of the four corner points to obtain actual coordinates and actual corner point probabilities of the four corner points; and

[0011] A parking space detection step is performed based on the center point coordinates and the center point probability and the actual coordinates and actual corner point probability of the four corner points.

[0012] Optionally, in the coordinate correction step, for each of the four corner points, a region of a preset size is divided with the second coordinate position of the corner point as the center, and the first coordinate of the corner point is searched for in the region, and the second coordinate is replaced with the searched first coordinate.

[0013] Optionally, in the parking space detection step, a final probability of the parking space is calculated based on the center point probability and the actual corner point probability of the four corner points.

[0014] Optionally, in the parking space detection step, the center point probability is P0, the probabilities of the corrected corner points of the four corner points are P1, P2, P3, and P4 respectively, and the final probability P of the parking space is 0.6 * P0 + 0.1 * (P1 + P2 + P3 + P4).

[0015] Optionally, in the image processing step, for the parking lot image, a preset picture size is adopted, picture features are extracted through multiple convolution layers and down-sampling layers, low-layer edge features and high-layer semantic features are fused using cross-layer connection, and then the picture size is restored to the preset picture size through an up-sampling layer, and center point coordinates and center point probability, first coordinates and first corner point probability of four corner points, and offset amounts (X, Y, L) of the four corner points relative to the center point coordinates are predicted on an output layer, wherein X, Y, and L respectively represent X direction and Y direction and distance.

[0016] Optionally, the output layer includes 10 corner point feature maps of the preset picture size and 12 offset amount feature maps of the preset picture size, wherein the first 5 layers of the corner point feature maps represent idle parking spaces, the last 5 layers represent occupied parking spaces, each of the 5 layers represents features of a center point and four corner points, the idle parking spaces and the occupied parking spaces share the offset amount feature maps, each corner point has three offset amounts, which respectively represent X direction and Y direction and distance L, there are four corner points, and there are 12 layers of features.

[0017] The parking space detection system of one aspect of the present application comprises:

[0018] An image acquisition module is configured to acquire a parking lot image around a vehicle, which contains one or more idle and occupied parking spaces.

[0019] an image processing module, configured to input the parking lot image into a neural network model for image processing to obtain a center point coordinate of a parking space, a center point probability, first coordinates of four corner points, first corner point probabilities of the four corner points, and offsets of the four corner points relative to the center point coordinate;

[0020] a coordinate calculation module, configured to calculate second coordinates of the four corner points according to the center point coordinate and the offsets of the four corner points relative to the center point coordinate;

[0021] a coordinate correction module, configured to correct the second coordinates of the four corner points by using the first coordinates of the four corner points to obtain actual coordinates of the four corner points and actual corner point probabilities of the four corner points; and

[0022] a parking space detection module, configured to perform parking space detection based on the center point coordinate, the center point probability, and the actual coordinates of the four corner points and the actual corner point probabilities of the four corner points.

[0023] Optionally, in the coordinate correction module, for each of the four corner points, a region of a preset size is divided with the second coordinate of the corner point as the center, and the first coordinate of the corner point is searched for in the region, and the second coordinate is replaced by the searched first coordinate.

[0024] Optionally, in the parking space detection module, a final probability of a parking space is calculated based on the center point probability and the actual corner point probabilities of the four corner points.

[0025] Optionally, in the parking space detection module, the center point probability is P0, the corrected corner points of the four corner points are P1, P2, P3, and P4 respectively, and the final probability P of the parking space is 0.6 * P0 + 0.1 * (P1 + P2 + P3 + P4).

[0026] Optionally, in the image processing module, for the parking lot image, a preset picture size is adopted, picture features are extracted through multiple convolution layers and down-sampling layers, edge features of low layers and semantic features of high layers are fused by using cross-layer connection, and then the picture features are restored to the preset picture size through up-sampling layers, and the center point coordinate of the parking space, the center point probability, the first coordinates of the four corner points, the first corner point probabilities of the four corner points, and offsets (X, Y, L) of the four corner points relative to the center point coordinate are predicted on an output layer, where X, Y, and L respectively represent an X direction, a Y direction, and a distance.

[0027] Optionally, the output layer includes 10 layers of corner point feature maps of the preset picture size and 12 layers of offset feature maps of the preset picture size, wherein the first 5 layers of the corner point feature maps represent idle parking spaces, the last 5 layers represent occupied parking spaces, each of the 5 layers represents features of a center point and 4 corner points, the idle parking spaces and the occupied parking spaces share the offset feature maps, each corner point has 3 offsets, respectively representing an X direction, a Y direction and a distance L, and there are 4 corner points, and the 12 layers of features.

[0028] The computer readable medium of one aspect of the present application has a computer program stored thereon, wherein the computer program is executed by a processor to implement the parking space detection method.

[0029] The computer device of one aspect of the present application comprises a storage module, a processor and a computer program stored on the storage module and executable on the processor, wherein the processor executes the computer program to implement the parking space detection method.

[0030] The parking space detection method and the parking space detection system according to the present application can effectively improve the parking space detection rate in a parking lot and improve the parking space positioning accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a flowchart of the parking space detection method of one aspect of the present application.

[0032] Figure 2 is a schematic diagram of a heat map for predicting a corner point.

[0033] Figure 3 is a schematic diagram showing the specific process of the image processing step.

[0034] Figure 4 is a schematic diagram of the coordinate calculation step.

[0035] Figure 5 is a schematic diagram of the coordinate correction step.

[0036] Figure 6 is a structural block diagram of the parking space detection system of one aspect of the present application. DETAILED DESCRIPTION

[0037] The following describes some of the multiple embodiments of the present application, which is intended to provide a basic understanding of the present application. It is not intended to identify key or decisive elements or limit the scope of protection.

[0038] For the sake of brevity and clarity, the principles of the application are described herein with reference to exemplary embodiments thereof. However, it will be readily apparent to those skilled in the art that the same principles can be equally applied to all types of parking space detection methods and parking space detection systems, and can be implemented therein, and that any such variations do not depart from the true spirit and scope of the present patent application.

[0039] Moreover, in the following description, reference is made to the accompanying drawings, which show, by way of example, specific embodiments thereof. Changes in form and detail can be made to these embodiments without departing from the spirit and scope of the application. Furthermore, while features of the application are disclosed with reference to only one of several embodiments / embodiments, this feature can be combined with one or more other features of other embodiments / embodiments, as can be desired and / or advantageous for any given or particular function. Therefore, the description herein is not to be taken in a limiting sense, and the scope of the present application is defined by the appended claims and their equivalents.

[0040] The terms such as "have", "include" and "comprise" indicate that the technical solutions of the present application do not exclude other elements (modules) and steps not directly or explicitly described in the specification and claims.

[0041] The concept of the parking space detection method of the present application is that, according to the parking lot picture containing one or more idle and occupied parking spaces around the vehicle obtained from the vehicle body camera, the picture is sent to the neural network for processing to obtain the center point coordinates and probability (P, X, Y) and the fine coordinates and probability of the four corner points of the parking space, and at the same time obtain the coordinate offset (direction X, Y and distance L) of the four corner points relative to the center point. According to the obtained center point coordinates and probability, all idle and occupied parking space center point coordinates are screened out. According to the coordinate offset of the four corner points relative to the center point of the parking space, the four corner point rough coordinates are calculated, and then in a certain range around each corner point, the fine coordinates of the four corner points obtained before are matched as the actual coordinates and probability of the four corner points of the parking space, and according to the center point coordinate probability and the four corner point actual coordinates and probability, the probability of the current parking space is recalculated for subsequent repeated parking space filtering logic.

[0042] Figure 1 is a flowchart of the parking space detection method of an aspect of the present application.

[0043] As Figure 1 shown, the parking space detection method of the present application includes the following steps:

[0044] Image acquisition step S100: acquiring a parking lot image containing one or more idle and occupied parking spaces around the vehicle;

[0045] Image processing step S200: input the parking lot image into the neural network model for image processing, to obtain the center point coordinates of the parking space, the first coordinates of the four corner points and the first corner point probability, and the offset of the four corner points relative to the center point coordinates;

[0046] Coordinate calculation step S300: calculate the second coordinates of the four corner points according to the center point coordinates and the offset of the four corner points relative to the center point coordinates;

[0047] Coordinate correction step S400: for the second coordinates of the four corner points, use the first coordinates of the four corner points to correct, to obtain the actual coordinates of the four corner points and the actual corner point probability; and

[0048] Parking space detection step S500: based on the center point coordinates and the center point probability and the actual coordinates of the four corner points and the actual corner point probability, to detect the parking space.

[0049] Next, steps S100-S500 are described in detail.

[0050] In the image acquisition step S100, the parking lot image containing one or more idle and occupied parking spaces around the vehicle can be acquired by the vehicle body camera.

[0051] In the image processing step S200, the parking space corner point coordinates are predicted based on the heat map, so that the first coordinates of the four corner points of the parking space and the first corner point probability, the center point coordinates of the parking space and the center point probability, and the offset of the four corner points relative to the center point coordinates can be obtained. Here, the "first coordinates" is also called "fine coordinates" because of its relatively high accuracy.

[0052] Figure 2 is a schematic diagram of a heat map for predicting corner points.

[0053] As Figure 2 shown, for each picture, the annotator will label all the parking space corner point coordinates in the picture. Since the labeling result usually has some errors, such as the actual coordinates and the labeled coordinates differ by a few pixels, if a neural network is used to predict the coordinates of this one pixel, only this pixel is a positive sample and other pixels are negative samples, the network can learn less features, and the result is often poor. In order to solve such problems, in the neural network model of the present application, for each parking space corner point, a series of point clouds are drawn within a certain range around it using Gaussian distribution as the prediction result of the model. By assigning a certain weight to the coordinates around the corner point instead of only predicting the peak value of one pixel of the corner point, the model can predict very accurate corner point coordinates.

[0054] Figure 3 is a schematic diagram representing the specific process of the image processing step.

[0055] As shown in Figure 3 , in the image processing step S200, for the input parking lot image, a picture size of 384*384 is adopted, the picture features are extracted through multiple convolution layers and down-sampling layers, the edge features of low layers and the semantic features of high layers are fused using cross-layer connection, and then the picture size is restored to 384*384 through up-sampling layers, the center point coordinates of the parking space, the center point probability, the first coordinates of the four corner points and the first corner point probability, and the offset amount (X, Y, L) of the four corner points relative to the center point coordinates are predicted on the output layer, wherein X, Y and L respectively represent the X direction and Y direction and the distance.

[0056] The output layer contains 10 layers of corner point feature maps of the preset picture size and 12 layers of offset amount feature maps of the preset picture size, wherein the first 5 layers of the corner point feature maps represent the idle parking spaces, the last 5 layers represent the occupied parking spaces, each of the 5 layers represents the features of the center point and the four corner points, the idle parking spaces and the occupied parking spaces share the offset amount feature maps, each corner point has 3 offset amounts, respectively representing the X direction and Y direction and the distance L, there are 4 corner points, and there are 12 layers of features.

[0057] Figure 4 is a schematic diagram of the coordinate calculation step.

[0058] Through the above-mentioned image processing step S200, the center point coordinates and the offset amount (X, Y, L) of the four corner points relative to the center point coordinates can be obtained. Then, in the coordinate calculation step S300, as shown in Figure 4 , the second coordinates of the four corner points (1~4) in Figure 3 can be calculated according to the center point coordinates (0) in Figure 3 and the offset amount (X, Y, L) of the four corner points relative to the center point coordinates, and here, the "second coordinates" are also called "rough coordinates" because their accuracy is not as accurate as the "fine coordinates" obtained by the heat map in step S200.

[0059] Figure 5 is a schematic diagram of the coordinate correction step.

[0060] As shown in Figure 5As shown, in the coordinate correction step S400, for the coarse coordinate of the corner point coordinate 1 of the four corner points calculated in the coordinate calculation step S300, a region of a preset size is circled at the coordinate position (i.e., a region of a preset size is divided with the second coordinate position of the corner point as the center), and the fine coordinate (i.e., the first coordinate obtained in the image processing step S200) of the corner point obtained from the heat map is searched in the region, and the coarse coordinate is replaced by the obtained fine coordinate, and similar processing is performed for the remaining three corner points, and finally the four corner points in the dashed box shown in the coordinate correction step S400, i.e., the corrected corner point coordinates, are obtained. Figure 4

[0061] In the parking space detection step S500, the center point coordinate and the center point probability and the actual coordinates of the four corner points are used for parking space detection. As an example, assuming that the center point probability is P0, the probabilities of the four corrected corner points are P1, P2, P3, and P4, respectively, and the final probability P of the parking space is 0.6 * P0 + 0.1 * (P1 + P2 + P3 + P4).

[0062] Due to various factors such as occlusion and unclear parking lines on the image, the error of directly regressing the four corner point offsets using the center point is large. If the corner points are predicted and grouped (such as using a clustering algorithm) as independent parking spaces, when there are multiple parking spaces in the image, the clustering result is often not stable and the post-processing is very complex, resulting in a large number of missed detections of parking spaces. In contrast, by using the parking space detection method of the present application, very accurate parking space coordinates can be obtained, and the recalculated parking space probability combines the probabilities of the center point and the four corner points, which can very accurately represent the probability that the parking space is a real parking space, and can greatly reduce the false detection and missed detection of parking spaces.

[0063] Figure 6 is a structural block diagram of a parking space detection system according to an aspect of the present application.

[0064] As shown in Figure 6 , the parking space detection system according to an aspect of the present application, characterized in that it comprises:

[0065] An image acquisition module 100 is configured to acquire a parking lot image around a vehicle, which contains one or more idle and occupied parking spaces.

[0066] An image processing module 200 is configured to input the parking lot image to a neural network model for image processing to obtain a center point coordinate and a center point probability of a parking space, first coordinates (fine coordinates) and first corner point probabilities of four corner points, and offsets of the four corner points relative to the center point coordinate.

[0067] ​The coordinate calculation module 300 is configured to calculate second coordinates (rough coordinates) of the four corner points according to the center point coordinates and offsets of the four corner points relative to the center point coordinates.

[0068] The coordinate correction module 400 is configured to correct the second coordinates of the four corner points by using the first coordinates of the four corner points to obtain actual coordinates of the four corner points and actual corner point probabilities.

[0069] The parking space detection module 500 is configured to perform parking space detection based on the center point coordinates, the center point probability, and the actual coordinates of the four corner points and the actual corner point probabilities.

[0070] In the coordinate correction module 400, for each of the second coordinates of the four corner points, a region of a specified size is circled at the coordinate position, and the first coordinates of the corner point are searched for in the region, and the second coordinates are replaced by the obtained first coordinates.

[0071] In the parking space detection module 500, the final probability of the parking space is calculated based on the center point probability and the actual corner point probabilities of the four corner points. As an example, in the parking space detection module 500, the center point probability is P0, the corrected corner points of the four corner points are P1, P2, P3, and P4, and the final probability P of the parking space is 0.6 * P0 + 0.1 * (P1 + P2 + P3 + P4).

[0072] In the image processing module 200, for the parking lot image, a preset picture size is used, picture features are extracted through multiple convolution layers and down-sampling layers, edge features of low layers and semantic features of high layers are fused by using cross-layer connection, and then the picture size is restored to the preset picture size through an up-sampling layer, and the center point coordinates and the center point probability of the parking space, the first coordinates and the first corner point probability of the four corner points, and the offsets (X, Y, L) of the four corner points relative to the center point coordinates are predicted on an output layer, where X, Y, and L respectively represent the X direction, the Y direction, and the distance.

[0073] The output layer includes 10 corner point feature maps of the preset picture size and 12 offset feature maps of the preset picture size, where the first 5 layers of the corner point feature maps represent idle parking spaces, the last 5 layers represent occupied parking spaces, each of the 5 layers represents features of a center point and four corner points, the idle parking spaces and the occupied parking spaces share the offset feature maps, each corner point has three offsets, which respectively represent the X direction, the Y direction, and the distance L, there are four corner points, and there are 12 layers of features.

[0074] The application further provides a computer readable medium, which stores a computer program, and the computer program is executed by a processor to implement the parking space detection method.

[0075] The application further provides a computer device, which comprises a storage module, a processor and a computer program stored in the storage module and executable on the processor, and the processor implements the parking space detection method when executing the computer program.

[0076] As described above, in the parking space detection method and the parking space detection system, a neural network model is designed, an image with a parking space is input, a feature map of four corner points of a parking space, a feature map of a center point of the parking space, a coordinate offset of the corner points relative to the center point, and a parking space state are output, thereby bringing the following technical effects: a lightweight network design is adopted, and real-time operation on an edge computing platform is realized; a parking space detection rate is high, a false detection rate is low, and the method can be used as a visual perception part of a full-automatic parking system and be combined with a radar signal; the method supports regular parking spaces such as horizontal and vertical parking spaces and irregular parking spaces such as diagonal brick surfaces, and the algorithm has strong robustness.

[0077] Moreover, the parking space detection method and the parking space detection system can effectively solve the following problems: a low parking space detection rate when only relying on a radar for automatic parking in a parking lot; a low positioning accuracy of a parking space when only using a center point and a corner point offset for detection in other parking space detection methods; and the method supports output of a parking space validity state, and outputs a parking space that cannot be parked when a car occupies the parking space or a cone barrel or a ground lock is used.

[0078] As described above, the parking space detection method and the parking space detection system can provide visual perception of a parking space for an automatic parking system, improve a parking space detection rate and a corner point positioning accuracy, and a deep learning network suitable for an edge computing platform is provided in the application, thereby realizing lightweight network design and a high parking space detection rate, and the application simultaneously supports detection of idle parking spaces and occupied parking spaces, can support display of the two kinds of parking spaces on a UI of an automatic parking system, and provides a better display experience.

[0079] The above examples mainly illustrate the parking space detection method and the parking space detection system. Although only some specific embodiments of the application are described, those skilled in the art should understand that the application can be implemented in many other forms without departing from the spirit and scope of the application. Therefore, the examples and embodiments are regarded as illustrative rather than limiting, and the application can cover various modifications and alternatives without departing from the spirit and scope of the application as defined in the appended claims.

Claims

1. A parking space detection method, characterized in that, include: The image acquisition step involves acquiring images of the parking lot surrounding the vehicle, including one or more vacant and occupied parking spaces. The image processing step involves inputting the parking lot image into a neural network model for image processing to obtain the center point coordinates and center point probability of the parking space, the first coordinates and first corner point probability of the four corner points, and the offset of the four corner points relative to the center point coordinates. The coordinate calculation step involves calculating the second coordinates of the four corner points based on the coordinates of the center point and the offsets of the four corner points relative to the center point coordinates. The coordinate correction step involves correcting the second coordinates of the four corner points using the first coordinates of the four corner points to obtain the actual coordinates and actual corner point probabilities of the four corner points. as well as The parking space detection step involves detecting parking spaces based on the coordinates and probability of the center point, as well as the actual coordinates and probabilities of the four corner points.

2. The parking space detection method as described in claim 1, characterized in that, In the coordinate correction step, for each of the four corner points, a region of a preset size is divided with the second coordinate position of the corner point as the center. The first coordinate of the corner point is found in the region, and the second coordinate is replaced with the found first coordinate.

3. The parking space detection method as described in claim 1, characterized in that, In the parking space detection step, the final probability of the parking space is calculated based on the center point probability and the actual corner point probabilities of the four corner points.

4. The parking space detection method as described in claim 3, characterized in that, In the parking space detection step, the probability of the center point is P0, the probabilities of the corrected corner points of the four corner points are P1, P2, P3, and P4, respectively, and the final probability P of the parking space is 0.6 * P0 + 0.1 * (P1 + P2 + P3 + P4).

5. The parking space detection method as described in claim 1, characterized in that, In the image processing step, for the parking lot image, a preset image size is used, and image features are extracted through multiple convolutional layers and downsampling layers. Cross-layer connections are used to fuse low-level edge features and high-level semantic features. Then, an upsampling layer is used to restore the preset image size. On the output layer, the center point coordinates and center point probability of the parking space, the first coordinates and first corner point probabilities of the four corner points, and the offsets (X, Y, L) of the four corner points relative to the center point coordinates are predicted, where X, Y, and L represent the X direction, Y direction, and distance, respectively.

6. The parking space detection method as described in claim 5, characterized in that, The output layer contains 10 layers of corner feature maps of the preset image size and 12 layers of offset feature maps of the preset image size. The first 5 layers of the corner feature maps represent vacant parking spaces, and the last 5 layers represent occupied parking spaces. Each of the 5 layers represents the features of the center point and 4 corner points. Vacant and occupied parking spaces share the offset feature map. Each corner point has 3 offsets, representing the X direction, the Y direction, and the distance L, respectively, for a total of 4 corner points and 12 layers of features.

7. A parking space detection system, characterized in that, include: The image acquisition module is used to acquire images of the parking lot surrounding the vehicle, including one or more vacant and occupied parking spaces. The image processing module is used to input the parking lot image into a neural network model for image processing to obtain the center point coordinates and center point probability of the parking space, the first coordinates and first corner point probability of the four corner points, and the offset of the four corner points relative to the center point coordinates. The coordinate calculation module is used to calculate the second coordinates of the four corner points based on the coordinates of the center point and the offsets of the four corner points relative to the coordinates of the center point. The coordinate correction module is used to correct the second coordinates of the four corner points using the first coordinates of the four corner points to obtain the actual coordinates and actual corner point probabilities of the four corner points. as well as The parking space detection module is used to detect parking spaces based on the coordinates and probability of the center point and the actual coordinates and probabilities of the four corner points.

8. The parking space detection system as described in claim 7, characterized in that, In the coordinate correction module, for each of the four corner points, a region of a preset size is divided with the second coordinate position of the corner point as the center. The first coordinate of the corner point is found in the region, and the second coordinate is replaced with the found first coordinate.

9. The parking space detection system as described in claim 7, characterized in that, In the parking space detection module, the final probability of the parking space is calculated based on the center point probability and the actual corner point probabilities of the four corner points.

10. The parking space detection system as described in claim 8, characterized in that, In the parking space detection module, the probability of the center point is P0, the corrected corner points of the four corner points are P1, P2, P3, and P4, and the final probability P of the parking space is 0.6 * P0 + 0.1 * (P1 + P2 + P3 + P4).

11. The parking space detection system as described in claim 7, characterized in that, In the image processing module, for the parking lot image, a preset image size is used, and image features are extracted through multiple convolutional layers and downsampling layers. Cross-layer connections are used to fuse low-level edge features and high-level semantic features. Then, an upsampling layer is used to restore the preset image size. On the output layer, the center point coordinates and center point probability of the parking space, the first coordinates and first corner point probabilities of the four corner points, and the offsets (X, Y, L) of the four corner points relative to the center point coordinates are predicted, where X, Y, and L represent the X direction, Y direction, and distance, respectively.

12. The parking space detection system as described in claim 11, characterized in that, The output layer contains 10 layers of corner feature maps of the preset image size and 12 layers of offset feature maps of the preset image size. The first 5 layers of the corner feature maps represent vacant parking spaces, and the last 5 layers represent occupied parking spaces. Each of the 5 layers represents the features of the center point and 4 corner points. Vacant and occupied parking spaces share the offset feature map. Each corner point has 3 offsets, representing the X direction, the Y direction, and the distance L, respectively, for a total of 4 corner points and 12 layers of features.

13. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the parking space detection method according to any one of claims 1 to 6.

14. A computer device comprising a storage module, a processor, and a computer program stored on the storage module and executable on the processor, characterized in that, When the processor executes the computer program, it implements the parking space detection method according to any one of claims 1 to 6.

Images