Parking space detection method, device, equipment, storage medium and vehicle

Abstract

The present disclosure relates to a parking space detection method, device, equipment, storage medium and vehicle, the method comprising: acquiring a parking space image; for each feature dimension related to the parking space, performing feature detection processing on the parking space image to obtain a parking space image feature group corresponding to the feature dimension, the parking space image feature group comprising a plurality of feature detection values corresponding to the feature dimension; for each parking space image feature group, performing parking space detection processing on the parking space image feature group corresponding to the feature dimension to obtain a detection result corresponding to the parking space image feature group, and the detection results corresponding to a plurality of parking space image feature groups forming a parking space detection result corresponding to the parking space image. The present disclosure performs multiple feature detection processing on the obtained parking space image for each feature dimension to obtain a plurality of parking space image feature groups, and then performs parking space detection according to the plurality of feature detection values included in each parking space image feature group to obtain a corresponding detection result, so that the parking space detection result is more accurate.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of image processing, and particularly relates to a parking space detection method and device, equipment, storage medium and vehicle. BACKGROUND

[0002] With the rapid development of the automobile industry, automatic driving is also paid more and more attention. Autonomous parking is an important function in automatic driving, and accurate detection of parking space information is needed to realize automatic parking.

[0003] In the related art, the data of each feature dimension obtained by identifying a certain parking space is only one by using the original collected image, thereby forming a set of parking space feature data, and the detection result of the parking space is determined based on the set of parking space feature data, so that the detection result is inaccurate. SUMMARY

[0004] To solve the above technical problems, the present disclosure provides a parking space detection method, device, equipment, storage medium and vehicle.

[0005] In a first aspect, the present disclosure provides a parking space detection method, comprising:

[0006] obtaining a parking space image;

[0007] performing feature detection processing on the parking space image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension, the parking space image feature group comprising a plurality of feature detection values corresponding to the feature dimension;

[0008] performing parking space detection processing on the parking space image feature group for the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group, and the detection results corresponding to a plurality of parking space image feature groups form a parking space detection result corresponding to the parking space image.

[0009] Optionally, before performing feature detection processing on the parking space image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension, the method further comprises:

[0010] performing down-sampling processing on the parking space image to obtain a parking space down-sampled image;

[0011] wherein performing feature detection processing on the parking space image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension comprises:

[0012] performing multiple times of feature detection processing on the parking space down-sampled image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension.

[0013] Optionally, the parking space image is down-sampled to obtain a parking space down-sampled image, including:

[0014] The parking space image is down-sampled at multiple preset scales to obtain a down-sampled feature map corresponding to each preset scale.

[0015] A down-sampled feature map corresponding to a target scale is selected from the down-sampled feature maps corresponding to the multiple preset scales.

[0016] The selected down-sampled feature map is subjected to image fusion processing to obtain a parking space down-sampled feature map.

[0017] Optionally, the selected down-sampled feature map is subjected to image fusion processing to obtain a parking space down-sampled feature map, including:

[0018] The selected down-sampled feature map is added to obtain a parking space down-sampled feature map while keeping the number of channels of the selected down-sampled feature map unchanged.

[0019] Optionally, for each feature dimension related to the parking space, the parking space image is subjected to feature detection processing to obtain a parking space image feature group corresponding to the feature dimension, including:

[0020] The parking space image is subjected to image analysis processing to obtain a feature map corresponding to each feature dimension.

[0021] For each feature dimension, the feature map corresponding to the feature dimension is subjected to multiple times of feature recognition processing to obtain a parking space image feature group corresponding to the feature dimension.

[0022] Optionally, the feature dimension includes at least two of the following:

[0023] A center point position dimension, and the parking space image feature group includes a center point position group.

[0024] A center point confidence dimension, and the parking space image feature group includes a center point confidence group.

[0025] An angle point regression value dimension, and the parking space image feature group includes an angle point regression value group.

[0026] A classification value dimension, and the parking space image feature group includes a classification value group.

[0027] Optionally, the parking space image is subjected to image analysis processing to obtain a feature map corresponding to each feature dimension, including:

[0028] The parking space image is input into a pre-trained parking space feature detection model for feature detection processing to obtain a feature map corresponding to each feature dimension output by the parking space feature detection model, and the parking space feature detection model is trained based on multiple sample parking space images and multi-dimensional feature true values corresponding to each sample parking space image.

[0029] Optionally, before the image analysis processing of the parking space image is performed to obtain the feature map corresponding to each feature dimension, the method further comprises:

[0030] Obtaining a plurality of sample parking space images;

[0031] For each sample parking space image, generating a multi-dimensional feature true value corresponding to the sample parking space image;

[0032] Based on the plurality of sample parking space images and the multi-dimensional feature true value corresponding to each sample parking space image, training a preset deep learning model to obtain a parking space feature detection model.

[0033] Optionally, the feature map comprises a center point confidence feature map, the feature dimension corresponding to the center point confidence feature map is a center point position dimension and / or a center point confidence dimension, and the multi-dimensional feature true value comprises a center point confidence true value;

[0034] The multi-dimensional feature true value corresponding to the sample parking space image comprises:

[0035] According to the parking space corner point position corresponding to the sample parking space image, calculating a center point coordinate of the parking space in the sample parking space image;

[0036] Setting a confidence corresponding to the center point coordinate as a first preset value and setting a confidence of other coordinates except the center point coordinate as a second preset value to obtain the center point confidence true value.

[0037] Optionally, the feature map comprises a classification value feature map, the feature dimension corresponding to the classification value feature map is a classification value dimension, and the multi-dimensional feature true value further comprises a classification value true value;

[0038] The multi-dimensional feature true value corresponding to the sample parking space image further comprises:

[0039] According to the center point coordinate, determining a redundant center point coordinate, and taking the center point coordinate and the redundant center point coordinate as parking space center point coordinates respectively;

[0040] Obtaining a sample classification value corresponding to the sample parking space image;

[0041] Setting a classification value corresponding to the parking space center point coordinate as the sample classification value and setting a classification value of other coordinates except the parking space center point coordinate as a third preset value to obtain the classification value true value.

[0042] Optionally, the feature map further comprises a corner point regression feature map, the feature dimension corresponding to the corner point regression feature map is a corner point regression value dimension, and the multi-dimensional feature true value further comprises a corner point regression true value;

[0043] The multi-dimensional feature true value corresponding to the sample parking space image further comprises:

[0044] According to the parking space corner point position, a sample regression value corresponding to each parking space midpoint coordinate is calculated;

[0045] The corner point regression value corresponding to each parking space midpoint coordinate is set as the sample regression value corresponding to the parking space midpoint coordinate, and the corner point regression value of other coordinates except the parking space midpoint coordinate is set as a fourth preset value, to obtain a corner point regression true value.

[0046] Optionally, the feature dimensions include a center point position dimension, a center point confidence degree dimension, a corner point regression value dimension, and a classification value dimension, and the feature map includes a center point confidence degree feature map, a classification value feature map, and a corner point regression feature map.

[0047] For each feature dimension, the feature map corresponding to the feature dimension is subjected to multiple feature recognition processes to obtain a parking space image feature group corresponding to the feature dimension, including:

[0048] The center point confidence degree feature map is subjected to center point recognition to obtain center point features of multiple candidate center points, and the center point features include candidate center point positions and candidate center point confidence degrees.

[0049] For each candidate center point, the corner point regression feature map is subjected to corner point recognition according to the center point position of the candidate center point feature to obtain a corner point regression value corresponding to the candidate center point.

[0050] For each candidate center point, the classification value feature map is subjected to classification recognition according to the center point position of the candidate center point feature to obtain a classification value corresponding to the candidate center point.

[0051] The multiple candidate center point positions form a center point position group, the multiple candidate center point confidence degrees form a center point confidence degree group, the multiple corner point regression values form a corner point regression value group, and the multiple classification values form a classification value group.

[0052] Optionally, before the center point confidence degree feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points, the method further includes:

[0053] The center point confidence degree feature map is subjected to non-local maximum value suppression processing to obtain a suppressed center point confidence degree feature map.

[0054] The center point confidence degree feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points, including:

[0055] The suppressed center point confidence degree feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points.

[0056] Optionally, the parking space image feature group is subjected to parking space detection processing of the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group, including:

[0057] The parking space image feature group is subjected to parking space detection processing of the corresponding feature dimension according to the mean value of the parking space image feature group to determine the detection result corresponding to the parking space image feature group.

[0058] Optionally, before the parking space image feature group is subjected to parking space detection processing of the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group for each parking space image feature group, the method further includes:

[0059] The feature analysis is performed on each parking space image feature group to obtain a feature analysis result corresponding to each parking space image feature group.

[0060] The parking space image feature group is subjected to parking space detection processing of the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group for each parking space image feature group, including:

[0061] If each feature analysis result does not satisfy a preset data mining condition corresponding to the feature analysis result, the parking space image feature group is subjected to parking space detection processing of the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group for each parking space image feature group.

[0062] Optionally, after the feature analysis is performed on each parking space image feature group to obtain a feature analysis result corresponding to each parking space image feature group, the method further includes:

[0063] If at least one feature analysis result satisfies a preset data mining condition corresponding to the feature analysis result, the multiple groups of parking space image feature groups are taken as mining data.

[0064] Optionally, the preset data mining condition corresponding to the feature analysis result includes:

[0065] The feature analysis result is greater than or equal to a preset threshold value corresponding to the parking space image feature group to which the feature analysis result belongs.

[0066] In a second aspect, the present disclosure provides a parking space detection device, including:

[0067] An acquisition module is configured to acquire a parking space image.

[0068] A detection module is configured to perform feature detection processing on the parking space image for each feature dimension related to a parking space to obtain a parking space image feature group corresponding to the feature dimension, the parking space image feature group including multiple feature detection values corresponding to the feature dimension.

[0069] The processing module is configured to perform, for each parking space image feature group, parking space detection processing on the parking space image feature group in a corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group, and the detection results corresponding to the plurality of parking space image feature groups form a parking space detection result corresponding to the parking space image.

[0070] In a third aspect, the present disclosure provides a computer device, comprising:

[0071] a memory;

[0072] a processor; and

[0073] a computer program;

[0074] The computer program is stored in the memory and is configured to be executed by the processor to implement the method of any one of the first aspect.

[0075] In a fourth aspect, the present disclosure provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the method of any one of the first aspect.

[0076] In a fifth aspect, the present disclosure provides a vehicle comprising the computer device of the third aspect.

[0077] The parking space detection method, device, equipment, storage medium and vehicle provided by the present disclosure obtain a parking space image, perform multiple feature detections on the image for each feature dimension related to a parking space, obtain a plurality of parking space image feature groups, so that each parking space image feature group includes a plurality of feature detection values, then perform corresponding parking space detection processing according to the plurality of feature detection values in each parking space image feature group to obtain detection results corresponding to different feature dimensions. Therefore, the image containing fewer parameters of a single parking space is detected multiple times in multiple dimensions, so that multiple data are detected for each feature dimension, a parking space image feature group corresponding to the feature dimension is formed, and the detection result of the corresponding feature dimension is obtained according to the plurality of feature detection value data included in each parking space image feature group, so that the detection result of the corresponding feature dimension is more accurate, and the parking space detection result corresponding to the parking space image formed by the detection results of the plurality of corresponding feature dimensions is also more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 A parking space detection method flowchart provided by an embodiment of the present disclosure;

[0079] Figure 2 An application scenario diagram provided by an embodiment of the present disclosure;

[0080] Figure 3 A down-sampling algorithm diagram provided by an embodiment of the present disclosure;

[0081] Figure 4 A training flowchart of a parking space detection model provided by an embodiment of the present disclosure is shown in FIG. 16;

[0082] Figure 5 An algorithm schematic diagram of a parking space detection model provided by an embodiment of the present disclosure is shown in FIG. 17;

[0083] Figure 6 A flowchart of generating a center point confidence true value provided by an embodiment of the present disclosure is shown in FIG. 18;

[0084] Figure 7 A flowchart of generating a classification value true value provided by an embodiment of the present disclosure is shown in FIG. 19;

[0085] Figure 8 A flowchart of generating an angle point regression true value provided by an embodiment of the present disclosure is shown in FIG. 20;

[0086] Figure 9 A multi-dimensional feature recognition processing method flowchart provided by an embodiment of the present disclosure is shown in FIG. 21;

[0087] Figure 10 A center point confidence feature map schematic diagram provided by an embodiment of the present disclosure is shown in FIG. 22;

[0088] Figure 11 An angle point regression feature map schematic diagram provided by an embodiment of the present disclosure is shown in FIG. 23;

[0089] Figure 12 A classification value feature map schematic diagram provided by an embodiment of the present disclosure is shown in FIG. 24;

[0090] Figure 13 A structure schematic diagram of a parking space detection device provided by an embodiment of the present disclosure is shown in FIG. 25;

[0091] Figure 14 A structure schematic diagram of a computer device provided by an embodiment of the present disclosure is shown in FIG. 26. DETAILED DESCRIPTION

[0092] In order to more clearly understand the above-mentioned purposes, features and advantages of the present disclosure, the schemes of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0093] In the following description, many specific details are set forth in order to provide a thorough understanding of the present disclosure, but the present disclosure can also be implemented in other ways different from those described herein; obviously, the embodiments in the description are only some of the embodiments of the present disclosure, not all the embodiments.

[0094] In order to realize autonomous parking, an autonomous vehicle needs to accurately locate a parking position, and thus needs to detect a parking space in a surrounding environment to obtain accurate parking space information. A parking space corner point (referred to as a corner point) is an important feature of a parking space, which can represent the position of the parking space. In general, the edge line of a parking space is two parallel straight lines, and if the positions of four corner points are obtained, the complete position information of the parking space is essentially obtained because the connecting line of the corner points can represent the parking space line. Therefore, the corner points of the parking space are inherent attributes of the parking space and are suitable for most parking space scenarios. The types of parking spaces usually include, but are not limited to, non-character-type parking spaces, diagonal parking spaces, and linear parking spaces.

[0095] Currently, a method for detecting a parking space based on a regression manner usually outputs only one set of parking space corner points and one parking space confidence score for each parking space, and the parking space corner point regression and the parking space confidence score are not related, which cannot determine whether the corner point regression is reliable through the parking space confidence score, so that the detection result is inaccurate. To solve this problem, an embodiment of the present disclosure provides a parking space detection method, which will be introduced below in combination with specific embodiments.

[0096] Figure 1 A flowchart of a parking space detection method provided by an embodiment of the present disclosure is shown in FIG. 1. The method can be executed by a parking space detection device, which can be implemented in a software and / or hardware manner and can be configured in a computer device, such as a server or a terminal. The terminal specifically includes a vehicle machine or a vehicle-mounted device on a vehicle, etc. In addition, the method can be applied to the application scenario shown in FIG. 2, which includes a vehicle 201 and a parking space 202. It can be understood that the parking space detection method provided by an embodiment of the present disclosure can also be applied in other scenarios. Figure 2

[0097] The parking space detection method provided by an embodiment of the present disclosure will be introduced below in combination with the scenario shown in FIG. 2. For example, the vehicle machine 2011 in FIG. 2 can execute the method, which specifically includes the following steps. Figure 2 Figure 1 Figure 2

[0098] S101, acquiring a parking space image.

[0099] When the vehicle 201 is performing parking space detection, the vehicle machine 2011 will collect real-time images of the surrounding environment through the image collection device 2012 to obtain image data about the parking space. The image collection device 2012 can be, for example, a fisheye camera or other panoramic camera. The fisheye camera can independently realize wide-range and dead-angle-free monitoring, and collecting images about the parking space through the fisheye camera can avoid missing the parking space information in the collected images.

[0100] ​​​​S102, for each feature dimension related to the parking space, performing feature detection processing on the parking space image to obtain a parking space image feature group corresponding to the feature dimension, the parking space image feature group including a plurality of feature detection values corresponding to the feature dimension.

[0101] Exemplarily, the features related to the parking space 202 generally include a center point of the parking space 202, offset amounts of four corner points of the parking space 202 relative to the center point, and a category of the parking space 202. The vehicle machine 2011 will detect the image according to different dimensions of features to obtain a plurality of groups of data corresponding to each dimension of features. For example, the image is detected according to the center point of the parking space to obtain a center point feature group data, the image is detected according to the regression values of the four corner points relative to the candidate center point to obtain a corner point regression feature group data, and the image is detected according to the category of the parking space to obtain a category feature group data, etc. The center point feature group data includes a plurality of candidate center point detection values, the corner point regression feature group data includes the corresponding regression values of the four corner points relative to each candidate center point, and the category feature group data includes a plurality of category detection values.

[0102] S103, for each parking space image feature group, performing parking space detection processing on the parking space image feature group according to the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group, and the detection results corresponding to a plurality of parking space image feature groups form a parking space detection result corresponding to the parking space image.

[0103] Exemplarily, the parking space detection result generally includes a parking space position confidence, positions of four corner points of the parking space, and a category of the parking space. The vehicle machine 2011 can perform parking space position confidence detection according to the center point feature group data, obtain a confidence value of each candidate center point as the center point of the parking space based on a plurality of candidate center point detection values, and obtain the confidence of the parking space 202 in the current parking space detection result according to a plurality of confidence values. The positions of the four corner points of the parking space 202 can be obtained according to a plurality of candidate center point detection values and the corresponding regression values of the four corner points relative to each candidate center point. The category of the parking space 202 corresponding to the parking space 202 can be obtained according to a plurality of category detection values.

[0104] The embodiment of the present disclosure obtains a parking space image, performs feature detection on the image multiple times for each feature dimension related to the parking space, obtains multiple parking space image feature groups, so that each parking space image feature group includes multiple feature detection values, and then performs corresponding parking space detection processing according to the multiple feature detection values in each parking space image feature group to obtain detection results corresponding to different feature dimensions. Therefore, multiple multi-dimensional detection processing can be performed on the image containing fewer parameters of a single parking space, multiple data are detected for each feature dimension, a parking space image feature group corresponding to the feature dimension is formed, and the detection result of the corresponding feature dimension is obtained according to the multiple feature detection value data included in each parking space image feature group, so that the detection result of the corresponding feature dimension is more accurate, and thus the parking space detection result corresponding to the parking space image composed of multiple detection results of the corresponding feature dimensions is also more accurate.

[0105] On the basis of the above embodiment, before performing feature detection processing on the parking space image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension, the method further includes: performing down-sampling processing on the parking space image to obtain a parking space down-sampled image; and performing feature detection processing on the parking space image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension, including: performing multiple feature detection processing on the parking space down-sampled image for each feature dimension related to the parking space to obtain a parking space image feature group corresponding to the feature dimension.

[0106] Figure 3 is a schematic diagram of a down-sampling algorithm provided by the embodiment of the present disclosure. On the basis of the above embodiment, the down-sampling processing is performed on the image to obtain a parking space down-sampled feature map, including: performing down-sampling processing on the image at multiple preset scales to obtain a down-sampled feature map corresponding to each preset scale; selecting a down-sampled feature map corresponding to a target scale from the multiple down-sampled feature maps corresponding to the multiple preset scales; and performing image fusion processing on the selected down-sampled feature map to obtain a parking space down-sampled feature map.

[0107] The preset scale is an image scale that can be obtained by performing down-sampling on the original image by a preset multiple of the convolutional neural network, and the target scale is an image scale obtained by down-sampling based on a target multiple. Specifically, the target multiple and the target scale can be a specified multiple and a specified scale for image fusion determined from the preset multiple and the preset scale as needed.

[0108] For example, when the original size of the image is w in width and h in height, 2 times down-sampling is performed on the original image to obtain a down-sampled feature map with a size of (w / 2, h / 2), then 2 times down-sampling is performed on the down-sampled feature map with a size of (w / 2, h / 2) to obtain a 4 times down-sampled feature map with a size of (w / 4, h / 4), and so on, and further 2 times down-sampling is performed on the 4 times down-sampled feature map to obtain an 8 times down-sampled feature map with a size of (w / 8, h / 8).

[0109] In many cases, fusing features of different scales is an important means to improve the performance of image segmentation. Low-level features have higher resolution and contain more position and detail information, but have lower semantics and more noise due to fewer convolutions. High-level features have stronger semantic information, but have low resolution and poor perception of details. How to efficiently fuse the two and take the advantages and discard the disadvantages is the key to improving the segmentation model.

[0110] In the image fusion processing of the selected down-sampled images, the smallest scale image and the second smallest scale image are first fused, specifically: the smallest scale image and the second smallest scale image are first fused through a feature fusion network to obtain a fused image, and then feature extraction is performed on the fused image to obtain a feature map with the same scale as the fused image. Then, the obtained feature map is fused with the next second smallest scale image, specifically: the obtained feature map and the next second smallest scale image are first fused through a feature fusion network to obtain a new fused image, and then feature extraction is performed on the new fused image to obtain a new feature map with the same scale as the new fused image. In this way, all selected down-sampled images are fused, and the final obtained feature map is taken as the parking space down-sampled image. The scale of the parking space down-sampled image is the largest scale in the selected down-sampled images.

[0111] In order to reduce the number of parameters that the neural network needs to learn and prevent the results output by the neural network from overfitting, the images related to the parking space can be subjected to down-sampling processing of a preset scale, and at the same time, in order to retain as many features related to the parking space as possible, the down-sampled feature maps can be fused to obtain down-sampled feature maps of reasonable scales.

[0112] In this case, the image fusion processing of the selected down-sampled feature maps is performed to obtain a parking space down-sampled feature map, which includes: adding the selected down-sampled feature maps while keeping the number of channels of the selected down-sampled feature maps unchanged to obtain a parking space down-sampled feature map.

[0113] In the image fusion processing of the selected down-sampled feature maps, the down-sampled feature maps with a low scale can be fused into the down-sampled feature maps with a higher scale. For example, the feature fusion is performed by the ADD manner, which combines two feature vectors into a complex vector by a parallel strategy, that is, the ADD manner is to add the feature maps, the channel number remains unchanged, and the information quantity of each dimension is increased, and the dimension for describing the image is not increased.

[0114] The following will be described in combination with Figure 3 The down-sampling of the parking space image is exemplarily introduced, in which CNN represents the convolution operation of the image by the convolutional neural network, the down-sampling is achieved by extracting the main features through the pooling layer of the convolutional neural network, and the compression of the feature map. ADD represents the feature fusion operation of the down-sampled feature map by the feature fusion network. In order to obtain the parking space feature map with a target scale of 4 times down-sampling, the image 301 is first subjected to 5 preset scale down-sampling, including 2, 4, 8, 16, and 32 times down-sampling, to obtain the corresponding down-sampled feature maps 303, 304, 305, 306, and 307. For example, when the size of the image 301 is (w, h), the 2 times down-sampled feature map 303 is obtained by performing 2 times down-sampling on the feature map 302, and the size of the down-sampled feature map 303 is (w / 2, h / 2). Correspondingly, the 4 times down-sampled feature map 304 is obtained by performing 2 times down-sampling on the 2 times down-sampled feature map 303, and the size of the down-sampled feature map 304 is (w / 4, h / 4). The 8 times down-sampled feature map 305 is obtained by performing 2 times down-sampling on the 4 times down-sampled feature map 304, and the size of the down-sampled feature map 305 is (w / 8, h / 8). In this way, the 16 times down-sampled feature map 306 with a size of (w / 16, h / 16) and the 32 times down-sampled feature map 307 with a size of (w / 32, h / 32) can be obtained.

[0115] Then, the 4 times down-sampled feature map, the 8 times down-sampled feature map, and the 32 times down-sampled feature map 307 are fused as the target scale image. The 32 times down-sampled feature map 307 is selected and fused with the 8 times down-sampled feature map 305 to obtain the 8 times fused down-sampled feature map 308 with a size of (w / 8, h / 8). Next, the 4 times down-sampled feature map 304 is selected and fused with the 8 times fused down-sampled feature map 308 to obtain the 4 times down-sampled feature map 309 with a size of (w / 8, h / 8), which is taken as the parking space down-sampled feature map for subsequent steps.

[0116] The disclosure embodiment obtains the corresponding down-sampling feature map by performing down-sampling processing on the parking space image in multiple preset scales, and selects the down-sampling feature map corresponding to the target scale to perform image fusion processing to obtain the parking space down-sampling feature map. The number of parameters that the neural network needs to learn can be reduced by down-sampling processing, preventing overfitting of the results output by the neural network, while retaining as many features with high relevance to the parking space as possible.

[0117] On the basis of the above-mentioned embodiments, for each feature dimension related to the parking space, the parking space image is subjected to feature detection processing to obtain a parking space image feature group corresponding to the feature dimension, including: performing image analysis processing on the parking space image to obtain a feature map corresponding to each feature dimension; for each feature dimension, the feature map corresponding to the feature dimension is subjected to multiple times of feature recognition processing respectively to obtain a parking space image feature group corresponding to the feature dimension.

[0118] In some embodiments, the feature dimension includes at least two of the following: a center point position dimension, the parking space image feature group including a center point position group; a center point confidence dimension, the parking space image feature group including a center point confidence group; an angle point regression value dimension, the parking space image feature group including an angle point regression value group; a classification value dimension, the parking space image feature group including a classification value group.

[0119] In the process of determining the parking space, the positions of the four corner points of the parking space can be determined according to the center point position and the angle point regression value corresponding to the center point position. As introduced in the foregoing, the position of the angle point indicates the position of the parking space, so whether the detection of the center point is accurate affects whether the entire parking space detection result is accurate. Therefore, the center point position detected in the detection process has a center point confidence, which also represents the confidence of the entire parking space detection result. Meanwhile, the classification value of the center point position detected has a class ID corresponding to the class of the parking space detection result. Thus, the four corner point positions of the parking space, the confidence of the parking space, and the class of the parking space can be obtained to form a complete parking space detection result.

[0120] In this case, the image analysis processing on the parking space image to obtain a feature map corresponding to each feature dimension includes: inputting the parking space image into a pre-trained parking space feature detection model to perform feature detection processing, to obtain a feature map corresponding to each feature dimension output by the parking space feature detection model, the parking space feature detection model being trained based on multiple sample parking space images and multi-dimensional feature true values corresponding to each sample parking space image.

[0121] Correspondingly, before the image analysis processing on the parking space image to obtain a feature map corresponding to each feature dimension, the disclosure embodiment further provides a method for training a parking space feature detection model, Figure 4 A training flowchart of the parking space detection model provided by the disclosure embodiment is shown in FIG. 2.Figure 5 An algorithm schematic diagram of a parking space detection model provided for an embodiment of the present disclosure. Based on the above embodiment, referring to Figure 4 , the method for training a parking space feature detection model specifically includes the following steps:

[0122] S401, acquire a plurality of sample parking space images.

[0123] Illustratively, similar to acquiring the to-be-detected parking space image, four-way fisheye two-dimensional image data in the entire parking process can be collected by an image collection vehicle, and the four-way real-time images can be processed by a ring-view splicing algorithm to obtain a ring-view splicing image in the parking process. When acquiring the sample parking space image, images in different scenes, different parking space types, and different time points in the parking process are collected as samples, which can improve the richness of the sample parking space image and thus improve the generalization ability of the subsequently trained parking space feature detection model.

[0124] S402, for each sample parking space image, generate a plurality of groups of multi-dimensional feature true values corresponding to the sample parking space image.

[0125] For each sample parking space image, the corresponding dimensional feature true value is generated according to the feature of different dimensions. For example, according to the center point feature of the parking space, the corresponding center point confidence true value can be generated, and the center point confidence can reflect the confidence of this parking space detection; according to the category feature of the parking space, the corresponding classification value true value can be generated, and the classification value corresponds to the category identification number (i.e. ID) of the parking space; for example, according to the corner point regression feature of the parking space, the corner point regression true value is generated, etc.

[0126] S403, based on the plurality of sample parking space images and the plurality of groups of multi-dimensional feature true values corresponding to each sample parking space image, train a preset deep learning model to obtain a parking space feature detection model.

[0127] After obtaining the sample parking space image and the plurality of groups of multi-dimensional feature true values corresponding to the sample parking space image, the sample parking space image and the plurality of groups of multi-dimensional feature true values are used as training data to train a preset deep learning model. The preset deep learning model can be a convolutional neural network model, for example. The model parameters are adjusted according to the loss function to obtain the parking space feature detection model.

[0128] When training the preset deep learning model, the model parameters also need to be adjusted according to the loss function to obtain a parking space feature detection model with higher performance. The parking space feature detection model is used to output different dimensional feature maps according to the input to-be-detected parking space image. Therefore, the overall loss function of the model can include: calculating the position error between the parking space center point predicted value and the parking space center point true value, adjusting the model so that the output center point confidence feature map is more accurate, and the loss function used is the first cross-entropy loss function; calculating the classification error between the parking space classification value predicted value and the parking space classification value true value, adjusting the model so that the output classification value is more accurate, and the loss function used is the second cross-entropy loss function; calculating the corner point regression error between the parking space corner point predicted value and the parking space corner point true value, adjusting the model so that the output corner point regression value is more accurate, and the loss function used is the third cross-entropy loss function. EntropyLoss function (Cross Entropy Loss); calculate the error between the classification prediction value and the classification true value, adjust the model to make the output classification value feature map more accurate, and use the second cross Entropy Loss function (Cross Entropy Loss); calculate the error between the classification prediction value and the classification true value, adjust the model to make the output classification value feature map more accurate, and use the second cross Entropy Loss function and second cross Entropy The loss functions are essentially the same, except that they are used to calculate the error of different model parameters. Different loss functions are used for different dimensional feature detection of the parking space feature detection model to calculate the error of different dimensional feature detection values, so as to adjust the corresponding model parameters to make the output result of the parking space feature detection model more accurate.

[0129] The embodiments of the present disclosure obtain a sample parking space image, generate true values of different dimensional features for the sample parking space image, and train a preset deep learning model based on the sample parking space image and a plurality of sets of multi-dimensional feature true values to obtain a parking space feature detection model. The parking space feature detection model can extract features for different dimensional features to generate a feature map of corresponding parking space features, so that the subsequent parking space detection result is more accurate.

[0130] In some embodiments, the feature map includes a center point confidence feature map, the corresponding feature dimension of the center point confidence feature map is a center point position dimension and / or a center point confidence dimension, and the multi-dimensional feature true value includes a center point confidence true value.

[0131] Figure 6 The flowchart for generating a center point confidence true value provided by the embodiments of the present disclosure is shown as Figure 6 The specific steps include the following steps:

[0132] S601, according to the center point position of the sample parking space image corresponding to the parking space angle point, calculate the center point coordinates of the parking space in the sample parking space image.

[0133] For example, after obtaining the sample parking space image, for example, the size of the sample parking space image is width (w), height (h), the annotator will label the four corner points of the parking space frame in the sample parking space image, and calculate the coordinates of the center point of the parking space according to the four corner points of the parking space frame, for example, (x, y).

[0134] S602, set the confidence corresponding to the center point coordinates to a first preset value, and set the confidence of other coordinates except the center point coordinates to a second preset value to obtain the center point confidence true value.

[0135] Exemplarily, the confidence value corresponding to the center point coordinate (x, y) is set as a first preset value, which can be 1, and the confidence value of the coordinates of the points other than the center point coordinate (x, y) is set as a second preset value, which can be 0.

[0136] Correspondingly, in the case where the feature map output by the parking space feature detection model is a four-fold down-sampling feature map, the size of the feature map output by the model is (w / 4, h / 4), and therefore the ground truth used for training the model needs to be made based on the sample parking space image after four-fold down-sampling. For details of the down-sampling process, please refer to FIG. 5. Figure 3 The center point coordinate also needs to be down-sampled by 4, and is rounded to (round[x / 4], (round[y / 4]). Similarly, the confidence value corresponding to the center point coordinate is set as 1 in the sample parking space image after four-fold down-sampling, and the confidence value of other points is 0, so that the center point confidence ground truth can be obtained. In this way, each point in the center point confidence feature map output by the model has a corresponding confidence value, which can reflect the confidence of each point as a center point. The higher the confidence value, the greater the probability that the point is a real parking space center point.

[0137] The embodiments of the present disclosure set the first preset value for the confidence of the center point coordinate and the second preset value for the confidence of the coordinates other than the center point coordinate by calculating the coordinates of the parking space center point, so as to obtain the center point confidence ground truth to support model training.

[0138] In some embodiments, the feature map further includes a classification value feature map, the multi-dimensional feature ground truth further includes a classification value ground truth, and the plurality of parking space image feature groups further include a classification value group.

[0139] Figure 7 The flowchart for generating the classification value ground truth provided by the embodiments of the present disclosure specifically includes the following steps:

[0140] S701, determining redundant midpoint coordinates according to the center point coordinates.

[0141] Taking the four-fold down-sampling feature map output by the parking space feature detection model as an example, in order to increase the redundancy of the model for errors, the points within the range of (round[x / 4]±deltaX, round[y / 4]±deltaY) around the center point coordinate (round[x / 4], (round[y / 4]) in S502 can be selected as the redundant midpoints, and the coordinates of the redundant midpoints are determined.

[0142] S702, taking the center point coordinates and the redundant midpoint coordinates as the parking space midpoint coordinates, respectively.

[0143] The coordinates of the center point (round[x / 4], round[y / 4]) and the coordinates of the redundant midpoint are taken as the coordinates of the parking space midpoint.

[0144] S703, a sample classification value corresponding to the sample parking space image is obtained.

[0145] After obtaining the sample parking space image, the annotator will determine the class ID of the parking space according to the parking space displayed in the sample parking space image, and label the classification value corresponding to the class ID on the sample parking space image. The classification value pre-labeled on the sample parking space image is the sample classification value. The sample classification value corresponding to the sample parking space image is obtained.

[0146] S704, the classification value corresponding to the parking space midpoint coordinates is set as the sample classification value, and the classification value of other coordinates except the parking space midpoint coordinates is set as a third preset value, to obtain a classification value true value.

[0147] According to the sample classification value, the classification value corresponding to the parking space midpoint coordinates on the four times down-sampling feature map is set as the sample classification value, and the classification value of other coordinates except the parking space midpoint coordinates is set as a third preset value. The third preset value can be 0, representing that the points cannot be classified according to the points. However, the classification value of all parking space midpoints is the sample classification value. In this way, in the case that the rough position detection of the model on the parking space has a slight error, the parking space can still be accurately classified, and thus the classification value true value is obtained.

[0148] The classification value corresponding to the point in the classification value feature map corresponds to the class ID of the parking space. In order to make the parking space detection result more accurate, similar redundant operations to identifying the center point confidence group are also needed when the classification value group is identified according to the classification value feature map. Specifically, according to the coordinates of the candidate center point A detected in the center point confidence feature map, the point corresponding to the coordinates in the classification value feature map is found, the classification value of the point is recorded, and the classification values of the points in a preset range around the point (i.e., the points corresponding to the coordinates of the candidate center point B in the classification value feature map) are also recorded. The multiple classification values are grouped into a classification value group. The preset range is also (round[x / 4]±deltaX, round[y / 4]±deltaY).

[0149] The embodiments of the present disclosure determine the redundant midpoint coordinates through the center point coordinates, set the classification values corresponding to the center point coordinates and the redundant midpoint coordinates as the sample classification value, set the classification values of the coordinates of other points as a third preset value, and obtain the classification value true value to support model training.

[0150] In some embodiments, the feature map further includes an angle point regression feature map, the multi-dimensional feature true value further includes an angle point regression true value, and the plurality of parking space image feature groups further include an angle point regression value group.

[0151] Figure 8 The flowchart for generating the angle point regression true value provided by the embodiments of the present disclosure specifically includes the following steps:

[0152] S801 determines the redundant midpoint coordinates according to the center point coordinates.

[0153] Specifically, the implementation process and principles of S801 are the same as those of S701, which will not be repeated here.

[0154] S802 takes the center point coordinates and the redundant midpoint coordinates as the parking space midpoint coordinates respectively.

[0155] Specifically, the implementation process and principles of S802 are the same as those of S702, which will not be repeated here.

[0156] S803 calculates the sample regression value corresponding to each parking space midpoint coordinate according to the parking space angle point position.

[0157] For example, according to the coordinates of the four previously labeled angle points and the parking space midpoint coordinates, eight regression true values of the four angle points and each parking space midpoint are calculated. If it is determined that the feature map output by the parking space feature detection model is a four-fold down-sampling feature map, the regression true values also need to be four-fold down-sampled. For example, when one of the angle point coordinates is (xi, yi), the two regression true values of the angle point relative to the center point are dx = xi / 4-round(x / 4) and dy = yi / 4-round(y / 4). The calculation methods used for the regression true values of the remaining three angle points relative to the center point are the same and will not be repeated here. For example, the coordinates of the redundant midpoint are (x_i, y_i). Then the two regression true values of the angle point (xi, yi) relative to the redundant midpoint are dx = xi / 4-round(x_i / 4) and dy = yi / 4-round(y_i / 4). The calculation methods used for the regression true values of the remaining three angle points relative to the redundant midpoint are the same and will not be repeated here. The eight regression true values calculated relative to each parking space midpoint are taken as the sample regression value corresponding to the parking space midpoint.

[0158] S804 sets the angle point regression value corresponding to each parking space midpoint coordinate as the sample regression value corresponding to the parking space midpoint coordinate, and sets the angle point regression value of other coordinates except the parking space midpoint coordinates as a fourth preset value to obtain the angle point regression true value.

[0159] The sample regression value corresponding to the parking space midpoint coordinates is set as the label of the angle point regression value corresponding to the parking space midpoint coordinates, and the label corresponding to the angle point regression value of other coordinates except the parking space midpoint coordinates is set as a fourth preset value. The fourth preset value can be 0, which represents that the angle point regression value cannot be determined according to these points, so that the positions of the four angle points cannot be determined. Thus, the angle point regression true value is obtained.

[0160] The value corresponding to the point in the corner point regression feature map is eight corner point regression values. The corner point regression value represents the offset value of the parking space corner point relative to the parking space center point. The eight corner point regression values corresponding to the point represent the offset values of the four corner points relative to the point as the center point. Because the corner points of the parking space are four, and the coordinates are represented by two dimensions of X and Y, the center point corresponds to eight corner point regression values.

[0161] When the corner point regression value group is identified according to the corner point regression feature map, similar to the redundant operation of identifying the center point confidence group, the points in the preset range (round[x / 4]±deltaX, round[y / 4]±deltaY) around the coordinates of the detected candidate center point A on the corner point regression feature map are taken as candidate center point B, a plurality of candidate center points are obtained, and the eight corner point regression values corresponding to each point in the plurality of candidate center points are recorded to form the corner point regression value group.

[0162] The center point coordinate is determined to determine the redundant midpoint coordinate, the corner point regression values corresponding to the center point coordinate and the redundant midpoint coordinate are set as the sample regression value, and the corner point regression values of the coordinates of other points are set as the fourth preset value to obtain the focus point regression true value, thereby supporting model training.

[0163] Figure 9 A multi-dimensional feature identification processing method flowchart is provided for the embodiments of the present disclosure, Figure 10 A schematic diagram of the center point confidence feature map is provided for the embodiments of the present disclosure, Figure 11 A schematic diagram of the corner point regression feature map is provided for the embodiments of the present disclosure, Figure 12 A schematic diagram of the classification value feature map is provided for the embodiments of the present disclosure.

[0164] In some embodiments, when the feature dimensions include the center point position dimension, the center point confidence dimension, the corner point regression value dimension, and the classification value dimension, the feature maps include the center point confidence feature map, the classification value feature map, and the corner point regression feature map, and in combination with Figure 9 、 Figure 10 、 Figure 11 and Figure 12 , for each feature dimension, the feature map corresponding to the feature dimension is processed multiple times to obtain the parking space image feature group corresponding to the feature dimension, which specifically includes the following steps:

[0165] S901, center point identification is performed on the center point confidence feature map to obtain the center point features of a plurality of candidate center points. The center point features include the candidate center point position and the candidate center point confidence.

[0166] For example, according to the position true value setting process introduced above, the position true value of the center point is 1, and the position true value of other points is 0. The position value of a point in the center point confidence feature map is between 0 and 1. The specific steps of determining the candidate center point according to the center point confidence feature map are as follows: traversing all points in the center point confidence feature map, screening out points greater than a preset threshold 1001, for example, the preset threshold can be 0.9, which is taken as a candidate center point A (1001), and the coordinates and position value of the candidate center point 1001 are recorded. Correspondingly, according to the coordinates of the candidate center point 1001, other points in the range of deltaX and deltaY around the candidate center point 1001 (for example, when the size of the parking space is [Wspot, Hspot], then deltaX and deltaY change in the range of [0, min(Wspot, Hspot) / 2]), the other points in the range are taken as candidate center points B (including 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009), and the coordinates and position value of the candidate center points 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009 are recorded, thereby determining a plurality of candidate center points. The coordinates of the candidate center points represent the positions of the candidate center points, and the position values of the candidate center points represent the confidence of the points as center points. The coordinates of the points 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009 are recorded as a center point position group, and the position values of the points 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009 are recorded as a center point confidence group. The candidate center point A and the candidate center point B are both candidate center points, and there is no essential difference between them, but different names are taken to distinguish them when introducing them.

[0167] S902, for each candidate center point, according to the center point position of the candidate center point feature, the corner point regression feature map is identified to obtain the corner point regression value corresponding to the candidate center point.

[0168] For example, for the candidate center point A (1001) in S901 and each point in the plurality of candidate center points B (including 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009), according to the coordinates, the corresponding position point in the corner point regression feature map is found, that is, Figure 5In the embodiment, the points 1101, 1102, 1103, 1104, 1105, 1106, 1107, 1108, and 1109 of the corner point regression feature map (wherein the coordinates of 1001 and 1101 are the same, the coordinates of 1002 and 1102 are the same, and so on). These points all have labels, for example, the point 1101, the labels corresponding to the point 1101 include: when the point 1101 is taken as the center point of the parking space, the four predicted corner points corresponding to the point 1101, the eight corner point regression values of the point 1101, and the labels of the other points are similar, which are not described herein. The eight corner point regression values corresponding to the points 1101, 1102, 1103, 1104, 1105, 1106, 1107, 1108, and 1109 are recorded as a corner point regression value group.

[0169] S903, for each candidate center point, according to the center point position of the candidate center point feature, the classification value feature map is classified and recognized to obtain the classification value corresponding to the candidate center point.

[0170] For example, for the candidate center point A (1001) in S901 and each point in the plurality of candidate center points B (including 1002, 1003, 1004, 1005, 1006, 1007, 1008, and 1009), the corresponding position point in the classification value feature map is found according to the coordinates, that is, Figure 6 In the embodiment, the points 1201, 1202, 1203, 1204, 1205, 1206, 1207, 1208, and 1209 of the classification value feature map (wherein the coordinates of 1001 and 1201 are the same, the coordinates of 1002 and 1202 are the same, and so on). These points all have corresponding classification values, and the classification values corresponding to the points 1201, 1202, 1203, 1204, 1205, 1206, 1207, 1208, and 1209 are recorded as a classification value group.

[0171] In the embodiment, the plurality of candidate center point confidence values form a center point position group, the plurality of candidate center point confidence values form a center point confidence group, the plurality of corner point regression values form a corner point regression value group, and the plurality of classification values form a classification value group.

[0172] Therefore, the coordinates corresponding to the plurality of candidate center points form a center point position group, the position values corresponding to the plurality of candidate center points form a center point confidence group, the corner point regression values corresponding to the plurality of candidate center points form a corner point regression value group, and the classification values corresponding to the plurality of candidate center points form a classification value group. The center point confidence group is used to determine the confidence in the parking space detection result in the subsequent steps, each column of four regression values in the corner point regression value group corresponds to each candidate center point position in the center point position group, which is used to determine the four corner point positions of the parking space, that is, the parking space position in the parking space detection result in the subsequent steps, and the classification value group is used to determine the parking space classification in the parking space detection result in the subsequent steps.

[0173] The embodiment of the present disclosure obtains multiple corresponding dimensional feature maps through multiple dimensional detection processing of the parking space image, and then performs multiple feature recognition processing according to the multiple corresponding dimensional feature maps, so that each feature group of different dimensions includes multiple corresponding feature data, and the accuracy of the corresponding dimensional parking space detection result obtained according to the multiple feature data in each dimensional feature group can be improved.

[0174] Optionally, on the basis of the above-mentioned embodiment, before the center point confidence feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points, the method further includes: performing non-local maximum value suppression processing on the center point confidence feature map to obtain a suppressed center point confidence feature map; wherein the center point confidence feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points, including: the suppressed center point confidence feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points.

[0175] Before the center point confidence feature map is subjected to center point recognition to obtain the center point features of the multiple candidate center points, the center point confidence feature map can be traversed first, and the center point confidence feature map is subjected to non-local maximum value suppression processing. For example, the NMS algorithm (Non-Maximum Suppression, non-maximum suppression) can be used to suppress non-local maximum values. The basic principle of the NMS algorithm is to search for local maximum values and suppress non-maximum values.

[0176] In essence, when the non-local maximum value suppression processing is performed, a preliminary screening is performed, and some points with low confidence values and points with the same confidence values in the center point confidence feature map are removed. According to the principle of identifying candidate center points, the higher the confidence value corresponding to a point is, the greater the possibility that the point is a center point is. Therefore, points with low confidence values will not be identified as candidate center points, and points with the same confidence values do not need to be repeatedly identified. In theory, the positions of the points in the image will be very close, so only one point in the local area with the same confidence value needs to be selected as a representative.

[0177] After the center point confidence feature map is obtained, the center point confidence feature map is traversed again. Since the center point confidence feature map has been subjected to the non-local maximum value suppression processing, when the center point confidence feature map is traversed, it is not necessary to identify points with the same confidence value or points with a low confidence value again, and therefore the process of identifying the candidate center points is more efficient. In the identification process, a point A with a value greater than a preset threshold is screened out as a candidate center point, and the coordinates of the candidate center point A and the corresponding position value are recorded. Meanwhile, other points B in a preset range around the candidate center point A are also screened out as candidate center points, and the coordinates of the multiple candidate center points B and the corresponding position values are recorded.

[0178] The embodiments of the present disclosure make the candidate center point identification more efficient and faster by performing the non-local maximum value suppression processing on the center point confidence feature map and then determining the candidate center points according to the center point confidence feature map after the suppression.

[0179] On the basis of the above-mentioned embodiments, the parking space image feature group is subjected to parking space detection processing in the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group, including: performing parking space detection processing in the feature dimension corresponding to the parking space image feature group according to the mean value of the parking space image feature group to determine the parking space detection result in the feature dimension corresponding to the parking space image feature group.

[0180] For example, the parking space image feature group includes a center point position group, a center point confidence group, an angle point regression value group and a classification value group. According to the parking space confidence corresponding to the multiple candidate center points (including 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009) in the center point confidence group determined in S901, the average value of the parking space confidence is calculated, and the average value of the parking space confidence is taken as the parking space confidence in the parking space detection result.

[0181] According to the coordinates of the multiple candidate center points (including 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009) in the center point position group determined in S901 and the eight angle point regression values corresponding to the coordinates of each candidate center point (including 1101, 1102, 1103, 1104, 1105, 1106, 1107, 1108, 1109) determined in S902, the coordinates of the four angle points corresponding to each candidate center point are calculated, and the coordinates of the four angle points corresponding to each candidate center point are taken as a group of angle point coordinates. Then, the average value of the multiple groups of angle point coordinates is calculated, and the average value of the angle point coordinates is taken as the parking space angle point coordinates in the parking space detection result.

[0182] According to the classification value corresponding to the coordinates of the plurality of candidate center points (including 1201, 1202, 1203, 1204, 1205, 1206, 1207, 1208, and 1209) in the classification value group determined in S903, the average classification value of the plurality of classification values is calculated, and the class ID corresponding to the average classification value is taken as the parking space class in the parking space detection result.

[0183] The embodiments of the present disclosure can determine the detection result of the corresponding feature dimension by calculating the average value of the plurality of feature data in the parking space image feature group corresponding to each dimension, so that the accuracy of the parking space detection result is higher.

[0184] In some embodiments, optionally, before the method further includes: performing feature analysis on each parking space image feature group respectively to obtain a feature analysis result corresponding to each parking space image feature group; and performing parking space detection processing on the parking space image feature group in the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group for each parking space image feature group, including: if each feature analysis result does not satisfy the preset data mining condition corresponding to the feature analysis result, performing parking space detection processing on the parking space image feature group in the corresponding feature dimension to obtain a detection result corresponding to the parking space image feature group for each parking space image feature group.

[0185] For example, the plurality of feature detection values included in each parking space image feature group are analyzed to obtain a feature analysis result corresponding to each parking space image feature group, and the feature analysis result can reflect whether the detection values in the corresponding feature group are reliable. Compared with the above embodiment of calculating the mean value of the parking space image feature group to determine the detection result, the variance of the parking space position confidence group is calculated according to the parking space position confidence corresponding to the plurality of candidate center points. The variance of the plurality of groups of corner point coordinates is calculated as the variance of the group of corner point regression values according to the obtained coordinates of the plurality of groups of corner points. The variance of the classification value group is calculated according to the classification values corresponding to the plurality of candidate center points. When at least one of the variance of the center point confidence group, the variance of the corner point regression value group, and the variance of the classification value group is greater than or equal to the corresponding preset variance threshold, it means that the detection data of the plurality of parking space image feature groups is unreliable. When the variance of each parking space image feature group is less than the preset threshold corresponding to the parking space image feature group, it means that the detection data of the plurality of parking space image feature groups is reliable, and the parking space detection in the corresponding feature dimension of the parking space image feature group can be performed according to the parking space image feature group to obtain the parking space detection result in the corresponding feature dimension of the parking space image feature group.

[0186] If the variance of each parking space image feature group is less than the preset threshold corresponding to the parking space image feature group, for each parking space image feature group, a parking space detection corresponding to the feature dimension of the parking space image feature group is performed according to the parking space image feature group, to obtain a parking space detection result corresponding to the feature dimension of the parking space image feature group.

[0187] The variance of each parking space image feature group is calculated to evaluate whether the data of the parking space image feature group is reliable. When the variance of each parking space image feature group is less than the preset threshold corresponding to the parking space image feature group, a corresponding parking space detection result is obtained according to the parking space image feature group, which can improve the accuracy of the output parking space detection result.

[0188] In some embodiments, after the feature analysis of each parking space image feature group is performed to obtain a corresponding feature analysis result, the method further includes: if at least one feature analysis result satisfies a preset data mining condition corresponding to the feature analysis result, the multiple groups of parking space image feature groups are taken as mining data.

[0189] According to the feature analysis result corresponding to each parking space image feature group, it is determined whether the feature analysis result satisfies the preset data mining condition. For example, if it is determined through the feature analysis result that the detection value in the corresponding feature group is unreliable, it is determined that the feature analysis result satisfies the preset data mining condition. As long as one of the multiple feature analysis results satisfies the preset data mining condition, that is, the detection value of one of the multiple feature groups is unreliable, all the data in the multiple parking space image feature groups are taken as mining data.

[0190] In some embodiments, the preset data mining condition corresponding to the feature analysis result includes: the feature analysis result is greater than or equal to the preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs.

[0191] For example, when at least one of the variance of the center point confidence group, the variance of the angle point regression value group, and the variance of the classification value group is greater than or equal to the corresponding preset variance threshold, it is determined that the detection data of the multiple parking space image feature groups is unreliable, and the detection data of the multiple parking space image feature groups is taken as a negative sample to feed back to the model to improve the performance of the model.

[0192] In the embodiments of the present disclosure, multiple groups of data are analyzed according to the feature detection of each dimension, and then the effectiveness detection of the feature of the dimension is performed according to the feature analysis result and the preset condition, so as to mine the parking space detection data according to the detection result, realize data mining, and facilitate subsequent adjustment of the model.

[0193] Figure 13A structural schematic diagram of a parking space detection apparatus provided by an embodiment of the present disclosure is shown in FIG. 13. The parking space detection apparatus 1300 can be a car machine as described in the above embodiments, or the parking space detection apparatus 1300 can be a component or assembly in the car machine. The parking space detection apparatus 1300 provided by an embodiment of the present disclosure can perform the processing flow provided by the parking space detection method embodiments, as shown in FIG. 13, including the following steps. Figure 13

[0194] The acquisition module 1301 is configured to acquire a parking space image.

[0195] The detection module 1302 is configured to perform multiple times of multi-dimensional feature detection processing on the parking space image in relation to the parking space, to obtain multiple parking space image feature groups. Each parking space image feature group corresponds to a feature dimension, and each parking space image feature group includes multiple feature detection values.

[0196] The processing module 1303 is configured to, for each parking space image feature group, perform parking space detection processing on the parking space image feature group in the feature dimension corresponding to the parking space image feature group, to obtain a parking space detection result in the feature dimension corresponding to the parking space image feature group.

[0197] An embodiment of the present disclosure acquires a parking space image, and performs multiple times of feature detection on the image in relation to each feature dimension, to obtain multiple parking space image feature groups, so that each parking space image feature group includes multiple feature detection values. Then, according to the multiple feature detection values in each parking space image feature group, corresponding parking space detection processing is performed, to obtain detection results corresponding to different feature dimensions. Therefore, by performing multiple times of multi-dimensional detection processing on an image including fewer parameters of a single parking space, multiple data are detected in each feature dimension, to form a parking space image feature group corresponding to the feature dimension, and according to the multiple feature detection values included in each parking space image feature group, a detection result corresponding to the feature dimension is obtained, so that the detection result corresponding to the feature dimension is more accurate, and thus the parking space detection result corresponding to the parking space image composed of multiple detection results corresponding to the feature dimensions is also more accurate.

[0198] Optionally, the parking space detection apparatus 1300 further includes a down-sampling module 1304, configured to perform down-sampling processing on the parking space image, to obtain a parking space down-sampled image, before performing feature detection processing on the parking space image in relation to each feature dimension, to obtain a parking space image feature group corresponding to the feature dimension. When performing feature detection processing on the parking space image in relation to each feature dimension, to obtain a parking space image feature group corresponding to the feature dimension, the detection module 1302 is specifically configured to perform multiple times of feature detection processing on the parking space down-sampled image in relation to each feature dimension, to obtain a parking space image feature group corresponding to the feature dimension.

[0199] ​Optionally, the downsampling module 1304 is specifically configured to perform multiple preset scale downsampling processing on the parking space image to obtain a downsampling feature map corresponding to each preset scale, select a downsampling feature map corresponding to a target scale from the downsampling feature maps corresponding to the multiple preset scales, and perform image fusion processing on the selected downsampling feature map to obtain the parking space downsampling feature map when the downsampling processing is performed on the parking space image to obtain the parking space downsampling image.

[0200] Optionally, the downsampling module 1304 is specifically configured to add the selected downsampling feature map to obtain the parking space downsampling feature map while keeping the number of channels of the selected downsampling feature map unchanged when the image fusion processing is performed on the selected downsampling feature map to obtain the parking space downsampling feature map.

[0201] Optionally, the detection module 1302 is specifically configured to perform image analysis processing on the parking space image to obtain a feature map corresponding to each feature dimension, and perform multiple feature recognition processing on the feature map corresponding to each feature dimension to obtain the parking space image feature group corresponding to the feature dimension when the feature detection processing is performed on the parking space image for each feature dimension corresponding to the parking space.

[0202] Optionally, the feature dimensions include at least two of the following: a center point position dimension, the parking space image feature group including a center point position group; a center point confidence dimension, the parking space image feature group including a center point confidence group; a corner point regression value dimension, the parking space image feature group including a corner point regression value group; and a classification value dimension, the parking space image feature group including a classification value group.

[0203] Optionally, the downsampling module 1304 is specifically configured to input the parking space image into a pre-trained parking space feature detection model to perform feature detection processing, and obtain a feature map corresponding to each feature dimension output by the parking space feature detection model, when the image analysis processing is performed on the parking space image to obtain the feature map corresponding to each feature dimension, the parking space feature detection model being trained based on multiple sample parking space images and multi-dimensional feature true values corresponding to each sample parking space image.

[0204] Optionally, the parking space detection apparatus 1300 further includes a model training module 1305 configured to obtain multiple sample parking space images before the image analysis processing is performed on the parking space image to obtain the feature map corresponding to each feature dimension, generate multi-dimensional feature true values corresponding to each sample parking space image for each sample parking space image, and train a preset deep learning model based on the multiple sample parking space images and the multi-dimensional feature true values corresponding to each sample parking space image to obtain the parking space feature detection model.

[0205] Optionally, the feature maps comprise a center point confidence feature map, the feature dimension corresponding to the center point confidence feature map is a center point position dimension and / or a center point confidence dimension, and the multi-dimensional feature ground truth comprises a center point confidence ground truth; the model training module 1305 is specifically configured to, when generating the multi-dimensional feature ground truth corresponding to the sample parking space image, calculate the center point coordinates of the parking space in the sample parking space image according to the parking space corner point positions corresponding to the sample parking space image; set the confidence of the center point coordinates as a first preset value, and set the confidence of other coordinates except the center point coordinates as a second preset value, to obtain the center point confidence ground truth.

[0206] Optionally, the feature maps comprise a classification value feature map, the feature dimension corresponding to the classification value feature map is a classification value dimension, and the multi-dimensional feature ground truth further comprises a classification value ground truth; the model training module 1305 is further configured to, when generating the multi-dimensional feature ground truth corresponding to the sample parking space image, determine redundant center point coordinates according to the center point coordinates, and take the center point coordinates and the redundant center point coordinates as parking space center point coordinates respectively; obtain a sample classification value corresponding to the sample parking space image; set the classification value corresponding to the parking space center point coordinates as the sample classification value, and set the classification value of other coordinates except the parking space center point coordinates as a third preset value, to obtain the classification value ground truth.

[0207] Optionally, the feature maps further comprise a corner point regression feature map, the feature dimension corresponding to the corner point regression feature map is a corner point regression value dimension, and the multi-dimensional feature ground truth further comprises a corner point regression ground truth; the model training module 1305 is further configured to, when generating the multi-dimensional feature ground truth corresponding to the sample parking space image, calculate a sample regression value corresponding to each parking space center point coordinate according to the parking space corner point positions; set the corner point regression value corresponding to each parking space center point coordinate as the sample regression value corresponding to the parking space center point coordinate, and set the corner point regression value of other coordinates except the parking space center point coordinates as a fourth preset value, to obtain the corner point regression ground truth.

[0208] Optionally, the feature dimensions include a center point position dimension, a center point confidence dimension, an angle point regression value dimension, and a classification value dimension, and the feature maps include a center point confidence feature map, a classification value feature map, and an angle point regression feature map; wherein the detection module 1302 is specifically configured to, when performing multiple feature recognition processes on the feature map corresponding to each feature dimension to obtain the parking space image feature group corresponding to the feature dimension, perform center point recognition on the center point confidence feature map to obtain center point features of multiple candidate center points, the center point features including candidate center point positions and candidate center point confidences; perform angle point recognition on the angle point regression feature map according to the center point position of the candidate center point feature of each candidate center point to obtain an angle point regression value corresponding to the candidate center point; and perform classification recognition on the classification value feature map according to the center point position of the candidate center point feature of each candidate center point to obtain a classification value corresponding to the candidate center point; wherein the multiple candidate center point positions form a center point position group, the multiple candidate center point confidences form a center point confidence group, the multiple angle point regression values form an angle point regression value group, and the multiple classification values form a classification value group.

[0209] Optionally, the detection module 1302 is further configured to, before performing the center point recognition on the center point confidence feature map to obtain the center point features of the multiple candidate center points, perform non-local maximum value suppression on the center point confidence feature map to obtain a suppressed center point confidence feature map; wherein the detection module 1302 is configured to, when performing the center point recognition on the center point confidence feature map to obtain the center point features of the multiple candidate center points, perform the center point recognition on the suppressed center point confidence feature map to obtain the center point features of the multiple candidate center points.

[0210] Optionally, the processing module 1303 is specifically configured to, when performing the parking space detection process of the corresponding feature dimension on the parking space image feature group to obtain the detection result corresponding to the parking space image feature group, perform the parking space detection process of the corresponding feature dimension on the parking space image feature group according to the mean value of the parking space image feature group to determine the detection result corresponding to the parking space image feature group.

[0211] Optionally, the processing module 1303 is further configured to, before performing the parking space detection process of the corresponding feature dimension on each parking space image feature group to obtain the detection result corresponding to the parking space image feature group, perform feature analysis on each parking space image feature group to obtain a feature analysis result corresponding to each parking space image feature group; wherein the processing module 1303 is configured to, when performing the parking space detection process of the corresponding feature dimension on each parking space image feature group to obtain the detection result corresponding to the parking space image feature group, perform the parking space detection process of the corresponding feature dimension on each parking space image feature group to obtain the detection result corresponding to the parking space image feature group when each feature analysis result does not satisfy a preset data mining condition corresponding to the feature analysis result.

[0212] Optionally, the processing module 1303 is further configured to, after obtaining the feature analysis result corresponding to each parking space image feature group respectively by performing feature analysis on each parking space image feature group, use the multiple groups of parking space image feature groups as mining data when at least one feature analysis result meets a preset data mining condition corresponding to the feature analysis result.

[0213] Optionally, the preset data mining condition corresponding to the feature analysis result comprises: the feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs.

[0214] Figure 13 The parking space image processing apparatus of the embodiments shown can be used to execute the technical solutions of the method embodiments described above, and have similar implementation principles and technical effects, which will not be described here again.

[0215] Figure 14 A structural schematic diagram of a computer device is provided for the embodiments of the present disclosure. The computer device can be the terminal of the above embodiments. The computer device provided by the embodiments of the present disclosure can execute the processing procedure provided by the parking space detection method embodiments, such as Figure 14 As shown, the computer device 1400 includes a memory 1401, a processor 1402, a computer program and a communication interface 1403; wherein the computer program is stored in the memory 1401 and is configured to be executed by the processor 1402 to execute the parking space detection method as described above.

[0216] In addition, the embodiments of the present disclosure also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method described in the above embodiments.

[0217] The embodiments of the present disclosure also provide a vehicle, which includes Figure 14 The computer device shown.

[0218] It should be noted that, in this document, relationship terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

Claims

1. A parking space detection method characterized by comprising: The method comprises: acquiring a parking space image; performing feature detection processing on the parking space image for each feature dimension related to a parking space to obtain a parking space image feature group corresponding to the feature dimension, the parking space image feature group comprising a plurality of feature detection values corresponding to the feature dimension; performing parking space detection processing on the parking space image feature group for each feature dimension to obtain a detection result corresponding to the parking space image feature group, and a plurality of detection results corresponding to the parking space image feature groups form a parking space detection result corresponding to the parking space image; the feature dimension comprises at least two of the following: a center point position dimension, the parking space image feature group comprising a center point position group; a center point confidence dimension, the parking space image feature group comprising a center point confidence group; an angle point regression value dimension, the parking space image feature group comprising an angle point regression value group; a classification value dimension, the parking space image feature group comprising a classification value group.

2. The method of claim 1, wherein, Before the feature detection processing on the parking space image for each feature dimension related to a parking space to obtain a parking space image feature group corresponding to the feature dimension, the method further comprises: performing down-sampling processing on the parking space image to obtain a parking space down-sampled image; wherein the feature detection processing on the parking space image for each feature dimension related to a parking space to obtain a parking space image feature group corresponding to the feature dimension comprises: performing multiple times of feature detection processing on the parking space down-sampled image for each feature dimension related to a parking space to obtain a parking space image feature group corresponding to the feature dimension.

3. The method of claim 2, wherein, The down-sampling processing on the parking space image to obtain a parking space down-sampled image comprises: performing down-sampling processing on the parking space image at a plurality of preset scales to obtain a down-sampled feature map corresponding to each preset scale; selecting a down-sampled feature map corresponding to a target scale from a plurality of down-sampled feature maps corresponding to the preset scales; performing image fusion processing on the selected down-sampled feature map to obtain the parking space down-sampled feature map.

4. The method of claim 3, wherein, The image fusion processing on the selected down-sampled feature map to obtain the parking space down-sampled feature map comprises: adding the selected down-sampled feature map to obtain the parking space down-sampled feature map while keeping the number of channels of the selected down-sampled feature map unchanged.

5. The method of claim 1, wherein, The feature detection processing on the parking space image for each feature dimension related to a parking space to obtain a parking space image feature group corresponding to the feature dimension comprises: performing image analysis processing on the parking space image to obtain a feature map corresponding to each feature dimension; performing multiple times of feature recognition processing on the feature map corresponding to each feature dimension respectively to obtain a parking space image feature group corresponding to the feature dimension.

6. The method of claim 5, wherein, The image analysis processing on the parking space image to obtain a feature map corresponding to each feature dimension comprises: The parking space image is input into a pre-trained parking space feature detection model for feature detection processing to obtain a feature map corresponding to each feature dimension output by the parking space feature detection model.

7. The method of claim 6, wherein, Before the image analysis processing of the parking space image is performed to obtain the feature map corresponding to each feature dimension, the method further comprises: obtaining a plurality of sample parking space images; for each sample parking space image, generating a multi-dimensional feature true value corresponding to the sample parking space image; training a preset deep learning model based on the plurality of sample parking space images and the multi-dimensional feature true value corresponding to each sample parking space image to obtain the parking space feature detection model.

8. The method of claim 7, wherein, The feature map includes a center point confidence feature map, the feature dimension corresponding to the center point confidence feature map is a center point position dimension and / or a center point confidence dimension, and the multi-dimensional feature true value includes a center point confidence true value; wherein the generation of the multi-dimensional feature true value corresponding to the sample parking space image comprises: calculating the center point coordinates of the parking space in the sample parking space image according to the parking space corner point positions corresponding to the sample parking space image; setting the confidence of the center point coordinates to a first preset value and setting the confidence of other coordinates except the center point coordinates to a second preset value to obtain the center point confidence true value.

9. The method of claim 8, wherein, The feature map includes a classification value feature map, the feature dimension corresponding to the classification value feature map is a classification value dimension, and the multi-dimensional feature true value further includes a classification value true value; wherein the generation of the multi-dimensional feature true value corresponding to the sample parking space image further comprises: determining redundant center point coordinates according to the center point coordinates, and taking the center point coordinates and the redundant center point coordinates as parking space center point coordinates, respectively; obtaining a sample classification value corresponding to the sample parking space image; setting the classification value corresponding to the parking space center point coordinates to the sample classification value and setting the classification value of other coordinates except the parking space center point coordinates to a third preset value to obtain the classification value true value.

10. The method of claim 9, wherein, The feature map further includes an angle point regression feature map, the feature dimension corresponding to the angle point regression feature map is an angle point regression value dimension, and the multi-dimensional feature true value further includes an angle point regression true value; wherein the generation of the multi-dimensional feature true value corresponding to the sample parking space image further comprises: calculating a sample regression value corresponding to each parking space center point coordinate according to the parking space corner point positions; setting the angle point regression value corresponding to each parking space center point coordinate to the sample regression value corresponding to the parking space center point coordinate and setting the angle point regression value of other coordinates except the parking space center point coordinates to a fourth preset value to obtain the angle point regression true value.

11. The method of claim 5, wherein, The feature dimensions include a center point position dimension, a center point confidence dimension, an angle point regression value dimension and a classification value dimension, and the feature map includes a center point confidence feature map, a classification value feature map and an angle point regression feature map. Specifically, for each feature dimension, performing multiple feature recognition processes on the feature map corresponding to that feature dimension to obtain the parking space image feature group corresponding to that feature dimension includes: The center point confidence feature map is used to identify the center point, and the center point features of multiple candidate center points are obtained. The center point features include the position and confidence of the candidate center points. For each candidate center point, corner point identification is performed on the corner point regression feature map based on the center point position of the candidate center point features to obtain the corner point regression value corresponding to the candidate center point; For each candidate center point, the classification value feature map is classified and identified according to the center point position of the candidate center point features to obtain the classification value corresponding to the candidate center point; Among them, multiple candidate center point locations form the center point location group, multiple candidate center point confidence scores form the center point confidence group, multiple corner point regression values ​​form the corner point regression value group, and multiple classification values ​​form the classification value group.

12. The method of claim 11, wherein, Before performing center point identification on the center point confidence feature map to obtain center point features of multiple candidate center points, the method further includes: The center point confidence feature map is subjected to nonlocal maximum suppression processing to obtain a suppressed center point confidence feature map; The step of performing center point identification on the center point confidence feature map to obtain center point features of multiple candidate center points includes: The suppressed center point confidence feature map is used to identify the center point, thereby obtaining the center point features of the multiple candidate center points.

13. The method of claim 1, wherein, The step of performing parking space detection processing on the parking space image feature group according to the corresponding feature dimension to obtain the detection result corresponding to the parking space image feature group includes: Based on the mean of the parking space image feature group, parking space detection processing is performed on the corresponding feature dimension to determine the detection result corresponding to the parking space image feature group.

14. The method of claim 1, wherein, Before performing parking space detection processing on each of the parking space image feature groups according to the corresponding feature dimension to obtain the detection result corresponding to the parking space image feature group, the method further includes: Perform feature analysis on each of the parking space image feature groups to obtain the feature analysis results corresponding to each parking space image feature group; Specifically, for each parking space image feature group, performing parking space detection processing on the corresponding feature dimension to obtain the detection result corresponding to the parking space image feature group includes: If none of the feature analysis results meet the preset data mining conditions corresponding to the feature analysis results, for each parking space image feature group, parking space detection processing of the corresponding feature dimension is performed on the parking space image feature group to obtain the detection result corresponding to the parking space image feature group.

15. The method of claim 14, wherein, After performing feature analysis on each of the parking space image feature groups to obtain the feature analysis result corresponding to each parking space image feature group, the method further includes: If at least one of the feature analysis results satisfies a preset data mining condition corresponding to the feature analysis result, a plurality of groups of the parking space image features are taken as mining data.

16. The method of claim 14, wherein, The preset data mining condition corresponding to the feature analysis result includes: The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs.

17. A parking space detection device, characterized by The preset data mining condition corresponding to the feature analysis result includes: The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs. The feature analysis result is greater than or equal to a preset threshold corresponding to the parking space image feature group to which the feature analysis result belongs.

18. A computer device comprising: a memory; a processor; and a computer program; 19. A computer readable storage medium having stored thereon a computer program, characterized in that, wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method of any one of claims 1-16.

20. A vehicle characterized by The computer program is executed by the processor to implement the method of any one of claims 1-16. The computer device of claim 18 is included.

Images