Method and device for identifying obstacle vehicle position of parking system and vehicle

Abstract

The invention discloses a parking system obstacle vehicle position identification method and device and a vehicle, and belongs to the technical field of parking systems, and the identification method comprises the following steps: obtaining image data of the surrounding environment of a vehicle; performing multi-task joint detection on the image data to obtain image detection data, the image detection data at least comprising vehicle detection data and vehicle grounding point detection data; and obtaining a first vehicle position according to the vehicle detection data and the vehicle grounding point detection data so as to identify the obstacle vehicle position. The overall position of the obstacle vehicle can be identified through the vehicle detection data, and the contact position of the tires of the obstacle vehicle and the ground can be identified through the vehicle grounding point detection data, so that the boundary coordinates of the vehicle are accurately corrected, and the position deviation caused by distortion is compensated; the position of the obstacle vehicle is obtained through the vehicle detection data and the vehicle grounding point detection data, the problem that a single detection mode is insufficient in precision is effectively solved, and therefore the accuracy of obstacle vehicle position recognition is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of parking systems, and in particular to a method and device for identifying the position of an obstacle vehicle in a parking system, and a vehicle. BACKGROUND

[0002] An automatic parking system is a sub-module of an advanced driver assistance system (ADAS) and is used to automatically complete the operation of parking in a position. The automatic parking system can be simply divided into two parts: the first part is to find a parking space, and the second part is to plan and control. Most of the time, a parking space cannot be parked because the parking space has been parked by an obstacle vehicle, the obstacle vehicle has parked in multiple parking spaces due to a deviation, or the obstacle vehicle has parked in a line. Therefore, it is necessary to accurately identify the position of the static obstacle vehicle. When a dynamic obstacle vehicle is encountered during parking, it will affect the planned parking route, thereby causing a collision.

[0003] In related technologies, an automatic parking system inputs four-way fisheye information collected by a camera into a deep learning detection network to output the position and motion information of an obstacle vehicle, so as to identify the obstacle vehicle. The deep learning detection scheme based on the camera has low requirements for hardware such as vehicle computing power, does not need to go through information fusion, and can directly output the position and motion state of the obstacle vehicle. However, the output position of the obstacle vehicle may have problems such as jitter, deviation, and abnormal transient shift, which leads to a low detection accuracy of the obstacle vehicle. SUMMARY

[0004] Embodiments of the present application provide a method and device for identifying the position of an obstacle vehicle in a parking system, and a vehicle, which aims to solve the problem of low detection accuracy of the position of an obstacle vehicle output by a deep learning detection scheme based on a camera.

[0005] In a first aspect, a method for identifying the position of an obstacle vehicle in a parking system is provided, and the method comprises the following steps: Obtaining image data of the surrounding environment of the vehicle; Performing multi-task joint detection on the image data to obtain image detection data, wherein the image detection data at least includes vehicle detection data and vehicle grounding point detection data; Obtaining a first vehicle position according to the vehicle detection data and the vehicle grounding point detection data, so as to identify the position of the obstacle vehicle.

[0006] In some embodiments, the image detection data further includes vehicle contour data; The method further comprises: Performing error detection on the first vehicle position, and correcting the first vehicle position according to the vehicle contour data to obtain a second vehicle position in the case where an error is detected.

[0007] In some embodiments, the step of performing multi-task joint detection on the image data to obtain the image detection data comprises: performing multi-task joint detection on the image data based on the multi-task joint detection model to obtain the image detection data.

[0008] In some embodiments, the multi-task joint detection model comprises a plurality of convolution layers of different scales, a vehicle detection model, a vehicle grounding point detection model, and a vehicle contour segmentation model. The step of performing multi-task joint detection on the image data based on the multi-task joint detection model to obtain the image detection data comprises: extracting image features of different scales of the image data through the convolution layers of different scales; allocating the image features to the vehicle detection model, the vehicle grounding point detection model, and the vehicle contour segmentation model according to the scales of the image features; detecting the image features through the vehicle detection model to obtain vehicle detection data, detecting the image features through the vehicle grounding point detection model to obtain vehicle grounding point detection data, and detecting the image features through the vehicle contour segmentation model to obtain vehicle contour data.

[0009] In some embodiments, the vehicle detection data comprises a first anchor point and a first bounding box. The step of detecting the image features through the vehicle detection model to obtain the vehicle detection data comprises: detecting the image features through the vehicle detection model to identify a category of the image features; in a case where the category of the image features is a vehicle, detecting the image features through the vehicle detection model to determine a position of the first anchor point and a first offset; obtaining the first bounding box according to the position of the first anchor point and the first offset.

[0010] In some embodiments, the vehicle grounding point detection data comprises a second anchor point, a second bounding box, and a grounding point attribution point, the grounding point attribution point being a center point of the first bounding box. The step of detecting the image features through the vehicle grounding point detection model to obtain the vehicle grounding point detection data comprises: detecting the image features through the vehicle grounding point detection model to identify a category of the image features; in a case where the category of the image features is a vehicle grounding point, detecting the image features through the vehicle grounding point detection model to determine a position of the second anchor point, a second offset, and an attribution point offset; obtaining the second bounding box according to the position of the second anchor point, the second offset, and the attribution point offset.

[0011] In some embodiments, the step of detecting the image features by the vehicle contour segmentation model to obtain the vehicle contour data comprises: detecting the image features by the vehicle contour segmentation model to identify the category of the image features; in a case where the category of the image features is a vehicle, labeling the vehicle contour data by a polygonal line.

[0012] In some embodiments, the error at least includes a vehicle category error and a grounding point position jump error; In a case where the error is detected, the step of correcting the first vehicle position according to the vehicle contour data to obtain the second vehicle position comprises: in a case where the vehicle category error is detected, correcting the vehicle category error according to the vehicle contour data, recalculating the vehicle detection data and the vehicle grounding point detection data, and obtaining the second vehicle position; in a case where the grounding point position jump error is detected, determining the second vehicle position according to the positional relationship between the vehicle contour data and the grounding point.

[0013] In a second aspect, the embodiments of the present application further provide an identification device for the position of an obstacle vehicle of a parking system, the identification device comprising: an image acquisition module configured to acquire image data of the surrounding environment of the ego vehicle; an image detection module configured to perform multi-task joint detection on the image data to obtain image detection data, the image detection data at least including vehicle detection data and vehicle grounding point detection data; a first position detection module configured to obtain a first vehicle position according to the vehicle detection data and the vehicle grounding point detection data.

[0014] In a third aspect, the embodiments of the present application further provide a vehicle applying the identification method in the first aspect or the identification device in the second aspect.

[0015] The present application performs multi-task joint detection on the image data to obtain image detection data, which includes vehicle detection data and vehicle grounding point detection data. The vehicle detection data can be used to identify the overall position of the obstacle vehicle, and the vehicle grounding point detection data can be used to identify the position where the tire of the obstacle vehicle contacts the ground. The vehicle boundary coordinates can be accurately corrected, the positional deviation caused by distortion can be compensated, and the position of the obstacle vehicle can be obtained through the vehicle detection data and the vehicle grounding point detection data, effectively solving the problem of insufficient accuracy of single detection mode, thereby improving the accuracy of the position identification of the obstacle vehicle. BRIEF DESCRIPTION OF DRAWINGS

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for identifying the location of an obstacle vehicle in a parking system according to an exemplary embodiment of this disclosure. Figure 2 This is another flowchart illustrating a method for identifying the location of an obstacle vehicle in a parking system according to an exemplary embodiment of this disclosure; Figure 3 This is a flowchart illustrating step S202 of a method for identifying the location of an obstacle vehicle in a parking system according to an exemplary embodiment of this disclosure. Figure 4 This is a model architecture diagram of a multi-task joint detection model for an obstacle vehicle position identification method in a parking system provided by an exemplary embodiment of this disclosure; Figure 5 This is a schematic diagram of the detection structure of a vehicle detection model for a method for identifying the location of an obstacle vehicle in a parking system, provided by an exemplary embodiment of this disclosure. Figure 6 This is a schematic diagram of the detection structure of a vehicle grounding point detection model for a method for identifying the location of an obstacle vehicle in a parking system according to an exemplary embodiment of this disclosure; Figure 7 This is a schematic diagram of the structure of an obstacle vehicle position identification device for a parking system provided in an exemplary embodiment of this disclosure.

[0018] Explanation of icon numbers: 101. Image acquisition module; 102. Image detection module; 103. First position detection module; 104. Second position detection module. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0022] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not exclude the applicability to or configuration to devices performing additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more of the stated conditions or values ​​may in practice be based on additional conditions or values ​​beyond those stated.

[0023] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0024] In a first aspect, embodiments of this application provide a method for identifying the location of vehicles obstructed by obstacles in a parking system, such as... Figure 1 As shown, the identification method includes the following steps: S101. Acquire image data of the environment surrounding the vehicle.

[0025] The vehicle is equipped with vision sensors at the front, rear, left, and right to obtain image data covering the surrounding environment and provide data support for subsequent steps.

[0026] S102. Perform multi-task joint detection on the image data to obtain image detection data.

[0027] Image detection data includes at least vehicle detection data and vehicle contact point detection data. Vehicle detection data contains information such as the coordinates of the bounding rectangle of the obstacle vehicle and the vehicle category confidence level, used to locate the overall range of the vehicle. Vehicle contact point detection data contains information such as the coordinates of the bounding rectangles of the obstacle vehicle's tire contact points, used to accurately locate the vehicle's actual parking position.

[0028] Multi-task joint detection is a detection method that uses the same deep learning model framework to simultaneously complete multiple related detection tasks, such as vehicle detection and vehicle grounding point detection, to obtain vehicle detection data and vehicle grounding point detection data, and achieves feature weight sharing.

[0029] The detection method that uses shared feature weights for multi-task joint detection avoids redundant calculations from multiple independent models, significantly saving computing power and storage resources on the device.

[0030] S103. Obtain the first vehicle position based on the vehicle detection data and the vehicle grounding point detection data to identify the position of the obstacle vehicle.

[0031] Vehicle detection data can identify the overall position of obstacle vehicles, while vehicle ground contact point detection data can identify the contact position between the obstacle vehicle's tires and the ground. This accurately corrects the vehicle boundary coordinates, compensating for positional deviations caused by distortion. By obtaining the position of obstacle vehicles through both vehicle detection data and vehicle ground contact point detection data, the problem of insufficient accuracy of a single detection method is effectively solved, thereby improving the accuracy of obstacle vehicle position recognition.

[0032] This application also provides another method for identifying the location of obstacles in a parking system, such as... Figure 2 As shown, the identification method includes the following steps: S201. Acquire image data of the environment surrounding the vehicle.

[0033] In one embodiment, the vehicle is equipped with four fisheye lenses to obtain 360° surround image data around the vehicle. The fisheye lens is a lens with ultra-wide-angle characteristics, which can solve the blind spot problem of conventional cameras and adapt to the needs of close-range and large-area environmental perception in parking scenarios.

[0034] Throughout the entire process of vehicle locating and parking, four fisheye lenses simultaneously and continuously acquire images of the surrounding environment, obtaining image data. Utilizing their wide-angle distortion characteristics, the fisheye lenses cover an area from 0.5 meters to tens of meters around the vehicle, providing comprehensive data support for subsequent inspections and preventing obstacles from being missed due to limited field of view.

[0035] S202. Perform multi-task joint detection on the image data to obtain image detection data.

[0036] Image detection data is obtained by performing multi-task joint detection on image data based on a multi-task joint detection model. The image detection data includes at least vehicle detection data, vehicle grounding point detection data, and vehicle contour data.

[0037] The multi-task joint detection model includes multiple convolutional layers of different scales, a vehicle detection model, a vehicle grounding point detection model, and a vehicle contour segmentation model.

[0038] The steps for constructing a multi-task joint detection model include: labeling the image dataset used for training, annotating the bounding rectangles of obstacles and vehicles, the visible tire contact points of the vehicles, and the segmentation masks of the vehicle outlines, where the tire contact points are distinguished into four classes: left front point, right front point, left rear point, and right rear point. The dataset labels are then converted to the format required for the modeling objective, and the multi-task joint detection model is constructed.

[0039] The multi-task joint detection model integrates convolutional layers, vehicle detection models, vehicle grounding point detection models, and vehicle contour segmentation models into one model to achieve end-to-end recognition. It reduces the required computing resources and the number of modules to be deployed by sharing features, thereby improving the vehicle's ability to dynamically avoid obstacles when searching for and parking in the parking lot.

[0040] like Figure 3 As shown, step S202 includes S2021-S2025.

[0041] S2021. Extract image features of different scales from image data through convolutional layers of different scales.

[0042] like Figure 4 As shown, the multi-task joint detection model constructs multiple convolutional layers of different scales based on the Feature Pyramid Network (FPN) to extract image features of different scales. For example, inputting image data into convolutional layers of different scales yields image features of different scales, P2-P7, with the scales in P2-P7 increasing from small to large.

[0043] S2022. Assign image features to the vehicle detection model, vehicle grounding point detection model, and vehicle contour segmentation model according to the scale of the image features.

[0044] Image features such as P2 and P3 have high precision and can be used for vehicle grounding point detection and contour segmentation tasks. Image features such as P3-P7 have a larger receptive field and can predict larger and closer objects, such as vehicles and people nearby. Therefore, image features of P3-P7 are input into the vehicle detection model for vehicle detection, while image features of P2 and P3 are input into the vehicle grounding point detection model and the vehicle contour segmentation model, respectively, for detecting vehicle grounding points and vehicle contours.

[0045] S2023. Vehicle detection data is obtained by detecting image features using a vehicle detection model.

[0046] Because the shape of vehicles in images obtained with fisheye lenses is distorted, and the distortion increases with the distance of the obstacle vehicle from the vehicle, distortion can occupy most of the image when the obstacle vehicle is within 0.5 meters of the vehicle. This results in significant changes in the scale of the obstacle vehicle, making conventional single-scale detection methods and most detection algorithms based on anchor frames unsuitable.

[0047] Therefore, the vehicle detection model employs an anchorless, multi-size detection method to obtain vehicle detection data. This data includes a first anchor point and a first detection box. The first anchor point is a reference coordinate point adaptively generated by the vehicle detection model based on image features, used to calculate the position reference of the first detection box. The first detection box is the bounding rectangle of the vehicle predicted by the vehicle detection model based on image features. Due to the use of an anchorless prediction method, it can adapt to the size of the vehicle, rather than using a fixed anchor box size. The first detection box includes the vehicle's coordinate information, used to characterize the overall position of the obstacle vehicle. Because the image data undergoes convolutional layers, resulting in image feature sampling of P3-P7 by 4 to 128 times, the receptive field of the vehicle detection model can cover most of the fisheye image, thereby improving detection accuracy.

[0048] The image features are detected by a vehicle detection model to identify the category of the image features; when the category of the image features is vehicle, the image features are detected by the vehicle detection model to determine the position and first offset of the first anchor point; and the first detection box is obtained based on the position and first offset of the first anchor point.

[0049] like Figure 5 As shown, the vehicle detection model calculates the probability value of a candidate region in the image features as a vehicle. If the probability value is greater than a preset threshold, the candidate region in the features is determined to be a vehicle, and the loss is calculated through a classification loss function.

[0050] The formula for calculating the loss using the classification loss function is shown below: ; Where N represents the number of samples at the first anchor point. , Indicates weight, Indicates category label, This represents the probability value.

[0051] For candidate regions in image features classified as vehicles, a first anchor point is generated within the candidate region, and a first offset is predicted. The first offset includes the distances of the first anchor point in all directions, namely, the upper offset t, the lower offset b, the left offset l, and the right offset r. The loss is calculated using a bounding box regression loss function. A first detection box is obtained using the position of the first anchor point and the upper offset t, lower offset b, left offset l, and right offset r.

[0052] The formula for calculating the loss using the rectangular bounding box regression loss function is shown below: ; Where N represents the number of samples at the Anchor points. , , , These represent the predicted offsets to the boundary, , , , These represent the ground truth offsets from the Anchor point to the boundary.

[0053] By adapting the vehicle detection model without anchor frames to the features of fisheye images, the incompatibility between preset anchor frames and distorted vehicle shapes is avoided, thereby improving the detection accuracy of the overall vehicle position.

[0054] S2024. Vehicle grounding point detection data is obtained by detecting image features through a vehicle grounding point detection model.

[0055] Since the position of the conventional rectangular detection box for obstacle vehicles is not accurate enough, the tire contact point is predicted and expanded into a rectangular box as the detection target. This transforms the tire contact point regression problem into a tire contact area detection problem, thereby improving the positional accuracy of the tire contact point.

[0056] The vehicle contact point detection model also employs a frameless, multi-size detection method to obtain vehicle contact point detection data. This data includes a second anchor point, a second detection frame, and a contact point assignment point, which is the center point of the first detection frame. The second anchor point is a reference coordinate point generated based on image features, used to calculate the position reference of the second detection frame. The second detection frame is an extended detection area for a single tire contact point predicted by the vehicle contact point detection model based on image features. The second detection frame contains the position of the tire contact point, using a region range instead of the coordinates of a single point to improve the tolerance and detection accuracy of the tire contact point.

[0057] The image features are detected by the vehicle grounding point detection model to identify the category of the image features. When the category of the image features is vehicle grounding point, the image features are detected by the vehicle grounding point detection model to determine the position of the second anchor point, the second offset, and the home point offset. The second detection box is obtained based on the position of the second anchor point, the second offset, and the home point offset.

[0058] like Figure 6 As shown, the four tire contact points of the obstacle vehicle are detected. The tire contact points are divided into four categories: left front, right front, left rear, and right rear. The category of the tire contact point is predicted for the candidate region of the image features. The probability value of the candidate region is predicted. If the probability value is greater than a preset threshold, the category of the tire contact point is determined. The loss is calculated by the contact point classification loss function.

[0059] The formula for calculating the grounding point classification loss function is as follows: ; Where N represents the number of samples at the Anchor points. , Indicates weight, Indicates category label, This represents the probability value.

[0060] For each candidate region identified as a tire contact point, a second anchor point is generated within the candidate region. For each second anchor point, a second offset and a home point offset are predicted simultaneously.

[0061] The second offset includes the distances of the second anchor point in all directions, namely the upper offset t, lower offset b, left offset l, and right offset r, and the loss is calculated using the grounding point detection box regression loss function. The second detection box is obtained using the position of the second anchor point and the upper offset t, lower offset b, left offset l, and right offset r.

[0062] The attribution point offset is the distance between the second anchor point and the attribution point of the grounding point in the x and y directions. It is used to establish the association between the tire grounding point and the vehicle. The attribution point of the grounding point is the center point of the first detection frame, and the loss is calculated through the regression loss function of the attribution point of the grounding point.

[0063] The formula for calculating the regression loss function of the grounding point detection box is as follows: ; The formula for calculating the grounding point attribution regression loss function is as follows: ; Where N represents the number of samples at the Anchor points. , , , These represent the predicted offsets of the anchor points from the bounding box boundaries, respectively. , , , These represent the ground truth offsets from the Anchor point to the bounding box boundary, respectively. , These represent the predicted offsets of the Anchor point from the home point in the x and y directions, respectively. , These represent the true offsets of the Anchor point from the home point in the x and y directions, respectively.

[0064] By replacing point detection of tire contact points with region detection, even if there are deviations in the position of manually annotated contact points during model training, the second detection box can still cover the real contact points, reducing the impact of annotation accuracy on the results.

[0065] S2025. Vehicle contour data is obtained by detecting image features through a vehicle contour segmentation model.

[0066] Because the shape of obstacles and vehicles in images obtained by fisheye cameras is distorted, the vehicle body is no longer a straight line. Therefore, the vehicle outline is segmented and labeled using polygonal lines.

[0067] Image features are detected and their categories are identified using a vehicle contour segmentation model. When the image feature category is vehicle, the vehicle contour data is labeled with polygon lines.

[0068] Images P2 and P3, without high-magnification downsampling, fully preserve local details such as vehicle edges, corners, and textures, forming the core feature foundation for accurate contour segmentation. A vehicle contour segmentation model predicts the category of candidate feature regions in the images, calculates the probability that a candidate region is a vehicle, and distinguishes between vehicle and non-vehicle regions. When the probability that a candidate region is a vehicle exceeds a preset threshold, the vehicle's contour line is determined and labeled using polygons, and the loss is calculated using a vehicle contour loss function.

[0069] The formula for calculating the vehicle profile loss function is as follows: ; Where N represents the number of feature point samples. , Indicates weight, Indicates category label, This represents the probability value.

[0070] S203. Obtain the first vehicle position based on the vehicle detection data and the vehicle grounding point detection data to identify the position of the obstacle vehicle.

[0071] The overall position of the obstacle vehicle is determined based on the first anchor point and the first detection frame. The position of the tire contact point is determined based on the second anchor point and the second detection frame. The positional relationship between the tire contact point and the overall position of the obstacle vehicle is determined based on the relationship between the contact point's location and the second anchor point. The first vehicle position is obtained to identify the position of the obstacle vehicle.

[0072] The tire contact point, as a relatively stable location, provides a rigid constraint on the position of the obstacle vehicle, accurately corrects the vehicle boundary coordinates, and compensates for the position deviation caused by distortion. By obtaining the position of the obstacle vehicle through vehicle detection data and vehicle contact point detection data, the problem of insufficient accuracy of a single detection method is effectively solved, thereby improving the accuracy of obstacle vehicle position recognition.

[0073] S204. Perform error detection on the position of the first vehicle. If an error is detected, correct the position of the first vehicle based on the vehicle contour data to obtain the position of the second vehicle.

[0074] The errors include at least vehicle category error and ground point position jump error. Vehicle category error is a classification error caused by image distortion, background interference, or occlusion, which misclassifies non-vehicle targets as vehicles or vice versa, resulting in invalid or significantly deviated first or second detection boxes. Ground point position jump error is an error caused by abrupt changes in the tire ground point coordinates of the same obstacle vehicle in consecutive frames of image data, exceeding the reasonable range of actual movement or distortion, leading to abnormal calculation of the first vehicle position.

[0075] If a vehicle category error is detected, it is corrected based on the vehicle contour data. The vehicle detection data and vehicle grounding point detection data are then recalculated to obtain the second vehicle position. Alternatively, if a vehicle category error occurs due to a misjudgment of the vehicle as a whole, the vehicle position is repositioned using the vehicle contour data, the vehicle category error is corrected, and the vehicle detection data and vehicle grounding point detection data are recalculated to obtain the second vehicle position.

[0076] If a grounding point position fluctuation error is detected, the second vehicle position is determined based on the relationship between the vehicle contour data and the grounding point position. The vehicle contour data is a rigid constraint on the physical shape of the vehicle body, and the actual position of the grounding point is constrained by the vehicle contour data. Therefore, by using the vehicle contour data as a reference, valid grounding points are selected or reasonable coordinates are calculated through interpolation, thereby eliminating the fluctuation error.

[0077] Vehicle contour data can be used to perform secondary corrections on the positions of vehicles near obstacles, improving the detection accuracy. Furthermore, the multi-task joint detection model directly obtains vehicle contour data without requiring additional sensors or independent computing modules, thus saving computing resources.

[0078] Secondly, embodiments of this application also provide a device for identifying the location of vehicles in an obstacle within a parking system, such as... Figure 7 As shown, the recognition device includes an image acquisition module 101, an image detection module 102, a first position detection module 103, and a second position detection module 104.

[0079] The image acquisition module 101 is used to acquire image data of the vehicle's surrounding environment. Four fisheye lenses are mounted around the vehicle to obtain 360° surround image data. Fisheye lenses are lenses with ultra-wide-angle characteristics, which can solve the blind spot problem of conventional cameras and adapt to the close-range, large-area environmental perception needs in parking scenarios. Throughout the entire process of vehicle search and parking, the four fisheye lenses simultaneously and continuously acquire images of the surrounding environment, obtaining image data. The fisheye lenses utilize their wide-angle distortion characteristics to cover an area of ​​0.5 meters to tens of meters around the vehicle, providing complete data support for subsequent detection and avoiding the omission of obstacles due to limited field of view.

[0080] Image detection module 102 is used to perform multi-task joint detection on image data to obtain image detection data, which includes at least vehicle detection data and vehicle ground contact point detection data. Based on a multi-task joint detection model, image data is obtained by performing multi-task joint detection on the image data. The image detection data includes at least vehicle detection data, vehicle ground contact point detection data, and vehicle contour data. The multi-task joint detection model integrates convolutional layers, vehicle detection models, vehicle ground contact point detection models, and vehicle contour segmentation models into a single model, achieving end-to-end recognition. It reduces the required computing resources and the number of modules needed by sharing features, thereby improving the vehicle's ability to dynamically avoid obstacles during parking and search for parking spaces.

[0081] The first position detection module 103 is used to obtain the first vehicle position based on vehicle detection data and vehicle contact point detection data. The tire contact point, as a relatively stable position point, provides a rigid constraint on the position of the obstacle vehicle, accurately corrects the vehicle boundary coordinates, and compensates for the position deviation caused by distortion. By obtaining the position of the obstacle vehicle through vehicle detection data and vehicle contact point detection data, the problem of insufficient accuracy of a single detection method is effectively solved, thereby improving the accuracy of obstacle vehicle position recognition.

[0082] The second position detection module 104 is used to detect errors in the position of the first vehicle. If an error is detected, the position of the first vehicle is corrected based on the vehicle contour data to obtain the second vehicle position. The vehicle contour data allows for secondary correction of the position of obstacle vehicles, improving the detection accuracy. Furthermore, the multi-task joint detection model directly obtains the vehicle contour data without requiring additional sensors or independent computing modules, saving computational resources.

[0083] Thirdly, embodiments of this application also provide a vehicle that utilizes the identification method of the first aspect or the identification device of the second aspect. By performing multi-task joint detection on image data, image detection data is obtained. This image detection data includes vehicle detection data and vehicle ground contact point detection data. The vehicle detection data can identify the overall position of the obstacle vehicle, while the vehicle ground contact point detection data can identify the contact position between the obstacle vehicle's tires and the ground. This accurately corrects the vehicle boundary coordinates, compensating for positional deviations caused by distortion. By obtaining the position of the obstacle vehicle through both the vehicle detection data and the vehicle ground contact point detection data, the problem of insufficient accuracy of a single detection method is effectively solved, thereby improving the accuracy of obstacle vehicle position recognition.

[0084] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0085] The foregoing has provided a detailed description of a method and apparatus for identifying the location of an obstacle vehicle in a parking system, as provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying the location of an obstacle vehicle in a parking system, characterized in that, The identification method includes the following steps: Acquire image data of the environment surrounding the vehicle; The image data is subjected to multi-task joint detection to obtain image detection data, which includes at least vehicle detection data and vehicle grounding point detection data. The first vehicle position is obtained based on the vehicle detection data and the vehicle grounding point detection data, so as to identify the position of the obstacle vehicle.

2. The identification method according to claim 1, characterized in that, The image detection data also includes vehicle outline data; The identification method further includes: Error detection is performed on the position of the first vehicle. If an error is detected, the position of the first vehicle is corrected based on the vehicle contour data to obtain the position of the second vehicle.

3. The identification method according to claim 2, characterized in that, The steps for performing multi-task joint detection on the image data to obtain image detection data include: The image data is subjected to multi-task joint detection based on a multi-task joint detection model to obtain image detection data.

4. The identification method according to claim 3, characterized in that, The multi-task joint detection model includes multiple convolutional layers of different scales, a vehicle detection model, a vehicle grounding point detection model, and a vehicle contour segmentation model. The steps for performing multi-task joint detection on the image data based on the multi-task joint detection model to obtain image detection data include: Image features at different scales are extracted from the image data using convolutional layers of different scales; The image features are assigned to the vehicle detection model, the vehicle grounding point detection model, and the vehicle contour segmentation model according to their scale. The vehicle detection data is obtained by detecting the image features using the vehicle detection model, the vehicle grounding point detection data is obtained by detecting the image features using the vehicle grounding point detection model, and the vehicle contour data is obtained by detecting the image features using the vehicle contour segmentation model.

5. The identification method according to claim 4, characterized in that, The vehicle detection data includes a first anchor point and a first detection frame; The steps for obtaining vehicle detection data by detecting the image features using the vehicle detection model include: The image features are detected using the vehicle detection model to identify the category of the image features; When the category of the image feature is vehicle, the image feature is detected by the vehicle detection model to determine the position of the first anchor point and the first offset. The first detection box is obtained based on the position of the first anchor point and the first offset.

6. The identification method according to claim 5, characterized in that, The vehicle grounding point detection data includes a second anchor point, a second detection frame, and a grounding point attribution point, wherein the grounding point attribution point is the center point of the first detection frame. The steps for obtaining vehicle grounding point detection data by detecting the image features using the vehicle grounding point detection model include: The image features are detected using the vehicle grounding point detection model to identify the category of the image features; When the category of the image feature is vehicle grounding point, the image feature is detected by the vehicle grounding point detection model to determine the position of the second anchor point, the second offset, and the attribution point offset; The second detection frame is obtained based on the position of the second anchor point, the second offset, and the offset of the belonging point.

7. The identification method according to claim 6, characterized in that, The steps of obtaining the vehicle contour data by detecting the image features using the vehicle contour segmentation model include: The image features are detected using the vehicle contour segmentation model to identify the category of the image features; When the category of the image feature is vehicle, the vehicle outline data is labeled with polygonal lines.

8. The identification method according to claim 7, characterized in that, The error includes at least vehicle category error and grounding point position fluctuation error; The step of correcting the position of the first vehicle based on the vehicle contour data to obtain the position of the second vehicle when an error is detected includes: If the vehicle category error is detected, the vehicle category error is corrected based on the vehicle contour data, and the vehicle detection data and the vehicle grounding point detection data are recalculated to obtain the second vehicle position; If a positional fluctuation error at the grounding point is detected, the position of the second vehicle is determined based on the positional relationship between the vehicle contour data and the grounding point.

9. A device for identifying the location of an obstacle vehicle in a parking system, characterized in that, The identification device includes: The image acquisition module is used to acquire image data of the environment surrounding the vehicle. The image detection module is used to perform multi-task joint detection on the image data to obtain image detection data, which includes at least vehicle detection data and vehicle grounding point detection data. The first position detection module is used to obtain the first vehicle position based on the vehicle detection data and the vehicle grounding point detection data.

10. A vehicle, characterized in that, The identification method described in any one of claims 1-8 or the identification device described in claim 9 may be used.

Images