Roadside parking state judgment method and system based on wheel detection
By using a deep learning model based on wheel detection to identify vehicle wheel images and calculate virtual ground contact points and occupancy ratios, the robustness and cost issues of parking space identification in existing technologies are solved, and efficient parking space occupancy status judgment is achieved.
Patent Information
- Application Number
- CN202511369800.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-24
- Publication Date
- 2026-01-27
AI Technical Summary
Existing technologies cannot achieve robust parking space occupancy status recognition in complex scenarios such as occupancy, angled parking, and low light conditions. Traditional methods are greatly affected by environmental interference and are costly.
A wheel-based detection method is adopted, which identifies vehicle wheel images through a wheel detection model, calculates virtual ground contact points and generates candidate contact point lines, calculates the occupancy ratio to determine the parking space status, and uses a deep learning network to train a model for feature extraction and analysis.
Achieving robust automatic recognition of parking space occupancy status in complex visual interference environments improves recognition accuracy and adaptability while reducing hardware deployment costs.
Smart Images
Figure CN121415367A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent traffic management technology, and in particular to a method and system for determining roadside parking status based on wheel detection. Background Technology
[0002] Automatic identification of parking space occupancy status is the core function of an intelligent parking management system. It uses computer vision technology to analyze real-time images of parking spaces in surveillance videos to determine whether a parking space is occupied by a vehicle.
[0003] Existing mainstream technologies suffer from three inherent flaws, severely limiting recognition accuracy and scene adaptability. First, current vehicle detection methods rely on the overlap between the vehicle's bounding box and the parking space area for judgment. However, in real-world scenarios, vehicles are often partially obscured by adjacent vehicles, greenery, or buildings, resulting in incomplete vehicle outlines and over-reliance on global vehicle features susceptible to environmental interference. Second, physical sensor methods, which deploy hardware sensors in parking spaces to detect metal objects or height changes, avoid visual interference but can only perceive single-point states and cannot construct the spatial projection relationship between the vehicle and the parking space, leading to geometric misjudgments. Furthermore, large-scale deployment is costly. Third, background modeling methods based on pixel changes are significantly affected by sudden changes in illumination, rain and snow reflections, and low nighttime illumination, and the apparent pixel features lack a fundamental description of the vehicle's rigid structure. In contrast, wheels, as the only structure that is always in contact with the ground and least likely to be completely obscured, possess illumination-invariant geometric features, and the line connecting the contact points of the two wheels accurately reflects the vehicle's wheelbase projection.
[0004] Existing technologies, due to their adherence to the mindset of overall vehicle body detection or direct environmental perception, cannot achieve robust recognition in complex scenarios such as occlusion, oblique parking, and low light conditions.
[0005] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention
[0006] This invention provides a method and system for determining the roadside parking status based on wheel detection, which can effectively solve the problems in the background art.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A method for determining the roadside parking status based on wheel detection, the method comprising:
[0009] Acquire the video frames to be detected that cover the target parking space, wherein the geometric area of the target parking space is defined by a set of pre-calibrated position coordinates;
[0010] The video frame to be detected is input into the wheel detection model, and forward reasoning is performed based on the wheel detection model to output an image of each identified vehicle wheel.
[0011] Calculate the center point coordinates of each vehicle wheel image, and determine the center point coordinates as the virtual ground contact point of the corresponding vehicle wheel;
[0012] Traverse all the virtual ground contact points, combine the virtual ground contact points in pairs to generate a set of candidate landing point lines representing potential vehicle locations, and calculate the occupancy ratio value for each candidate landing point line.
[0013] Each of the occupancy ratio values is compared with a preset parking status determination threshold. If at least one of the occupancy ratio values is greater than the parking status determination threshold, the target parking space is determined to be occupied; otherwise, the target parking space is determined to be unoccupied.
[0014] Furthermore, the geometric region of the target parking space is defined by the coordinates of four vertices, which are labeled and recorded based on the image coordinate system to determine the precise coordinates of the four vertices of the target parking space.
[0015] Furthermore, the pixel data of the video frame to be detected is input into the input layer of the wheel detection model, and the wheel detection model performs layer-by-layer feature extraction and analysis to generate the position coordinates and confidence score of the bounding box of each vehicle wheel image.
[0016] Furthermore, the wheel detection model includes:
[0017] Collect and construct a training sample set of video frames including images of the vehicle wheels;
[0018] For each frame of the training sample set, the precise location of all wheel targets is marked, and corresponding label data is generated;
[0019] The training sample set with the labeled data is input into the wheel detection deep learning network architecture;
[0020] The weight parameters of the deep learning network architecture for wheel detection are iteratively adjusted using the backpropagation algorithm and gradient descent optimizer until the loss function of the wheel detection model on the validation set converges to a preset range, thus obtaining the wheel detection model.
[0021] Furthermore, the bounding box generated when the vehicle wheel image is recognized is a rectangle, and the position coordinates are defined by the center of the coordinates of the upper left corner vertex and the lower right corner vertex of the rectangle.
[0022] Furthermore, the candidate landing point line is a straight line segment connecting any two different virtual ground contact points. In a physical sense, the straight line segment represents the projection line of the two wheels of a single vehicle on the ground. The orientation and length of the straight line segment characterize the vehicle's posture in the parking scenario.
[0023] Furthermore, based on the length of each candidate landing point line, the length of the line segment that falls into the target parking space is compared and obtained.
[0024] Based on the length of the line segment falling into the target parking space and the total length of the target parking space, the corresponding occupancy ratio value is calculated for each of the candidate landing points.
[0025] Furthermore, the occupancy ratio value includes:
[0026] The formula for calculating the occupancy ratio is:
[0027] IOU ij =(l ij ∩S) / l ij (i,j = 1, 2, ..., m);
[0028] Among them, l ij The length of the line connecting the i-th and j-th virtual ground contact points represents the total length of the candidate landing points, and the area of the target parking space represents the total length of the line connecting the i-th and j-th virtual ground contact points. ij ∩S represents the length of the line segment that falls within the geometric region S, where IOU ij The calculated occupancy ratio value.
[0029] A roadside parking status determination system based on wheel detection, the system comprising:
[0030] The video acquisition module acquires video frames to be detected that cover the target parking space. The geometric area of the target parking space is defined by a set of pre-calibrated position coordinates.
[0031] The position recognition module inputs the video frame to be detected into the wheel detection model, performs forward reasoning based on the wheel detection model, and outputs the position information of each identified vehicle wheel image.
[0032] The center determination module calculates and extracts the coordinates of the center point of the bottom edge of the bounding box of each vehicle wheel image, and determines the center point coordinates as the virtual ground contact point corresponding to the vehicle wheel.
[0033] The occupancy calculation module iterates through all virtual ground contact points, combines the virtual ground contact points in pairs to generate a set of candidate landing point lines representing potential vehicle locations, and calculates the occupancy ratio value for each candidate landing point line.
[0034] The status judgment module compares each occupancy ratio value with a preset parking status judgment threshold. If at least one occupancy ratio value is greater than the parking status judgment threshold, the target parking space is determined to be occupied; otherwise, the target parking space is determined to be unoccupied.
[0035] Furthermore, the location recognition module includes:
[0036] The image training unit collects and constructs a training sample set of video frames, including images of vehicle wheels.
[0037] The label correspondence unit marks the precise location of all wheel targets for each frame of the image in the training sample set and generates corresponding label data;
[0038] The architecture generation unit inputs the training sample set with labeled data into the wheel detection deep learning network architecture;
[0039] The model generation unit iteratively adjusts the weight parameters of the training wheel detection deep learning network architecture based on the backpropagation algorithm and gradient descent optimizer until the loss function of the wheel detection model on the validation set converges to a preset range, thus obtaining the wheel detection model.
[0040] The technical solution of this invention can achieve the following technical effects:
[0041] It effectively solves the problem of how to achieve highly robust automatic recognition of parking space occupancy status in complex visual interference environments.
[0042] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0044] Figure 1This is a flowchart illustrating a method for determining the roadside parking status based on wheel detection.
[0045] Figure 2 This is a schematic diagram for wheel image recognition.
[0046] Figure 3 A schematic diagram for identifying the parking status of a vehicle. Detailed Implementation
[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0049] Example 1;
[0050] like Figure 1 As shown, this application provides a method for determining the roadside parking state based on wheel detection, including:
[0051] S10: Acquire the video frame to be detected covering the target parking space. The geometric area of the target parking space is defined by a set of pre-calibrated position coordinates.
[0052] S20: Input the video frame to be detected into the wheel detection model, perform forward reasoning based on the wheel detection model, and output the image of each identified vehicle wheel.
[0053] S30: Calculate the center point coordinates of each vehicle wheel image and determine the center point coordinates as the virtual ground contact point of the corresponding vehicle wheel;
[0054] S40: Traverse all virtual ground contact points, combine the virtual ground contact points in pairs to generate a set of candidate landing point lines representing potential vehicle locations, and calculate the occupancy ratio for each candidate landing point line.
[0055] S50: Compare each occupancy ratio value with a preset parking status determination threshold. If at least one occupancy ratio value is greater than the parking status determination threshold, the target parking space is determined to be occupied; otherwise, the target parking space is determined to be unoccupied.
[0056] Specifically, the process begins by acquiring video frames covering the target parking space. The geometric area of the target parking space is defined by multiple pre-calibrated coordinates, which can be done during camera installation to ensure that the camera's field of view accurately covers the parking space area, providing a precise reference background for subsequent image detection. During detection, the acquired video frames are input into a pre-trained wheel detection model. This model, based on deep learning technology and trained on a large number of samples from different environments and time periods, can quickly and accurately identify vehicle wheels in video frames under various complex lighting conditions. Through forward inference, the wheel detection model outputs an image of each identified vehicle wheel, providing basic data for subsequent calculations. Next, the center point coordinates of each vehicle wheel image are calculated as the virtual ground contact point of the corresponding wheel. Precise center point coordinate calculation effectively eliminates errors caused by... Interference from shooting angles or uneven lighting is mitigated to ensure the accuracy of virtual ground contact points. Then, all virtual ground contact points are traversed, and candidate landing point lines representing potential vehicle locations are generated by pairwise combinations. These lines simulate the possible actual positions of vehicles. The resulting lines visually reflect vehicle occupancy. An occupancy ratio is calculated for each candidate landing point line. Preferably, this ratio is determined by combining the actual size of the parking space and the distribution of the candidate lines within the geometric area. Finally, each occupancy ratio is compared to a preset parking status judgment threshold. This threshold is adjusted during system initialization to accommodate parking spaces of different sizes and usage scenarios. If at least one occupancy ratio exceeds the preset threshold, the target parking space is determined to be occupied; otherwise, it is determined to be unoccupied.
[0057] The technical solution of this invention effectively solves the problem of how to achieve highly robust automatic identification of parking space occupancy status in complex visual interference environments.
[0058] Furthermore, the geometric area of the target parking space is defined by the coordinates of its four vertices, which are labeled and recorded based on the image coordinate system, and the precise coordinates of the four vertices of the target parking space are recorded.
[0059] As a preferred embodiment of the above, firstly, during the camera installation or parking lot scene setup phase, technicians need to use precise measuring tools and methods to obtain the coordinates of the four vertices of the parking space and mark them in the image coordinate system. This process can be achieved through image calibration software, using a calibration board or other auxiliary tools to ensure the accuracy of the coordinates. Specifically, the selected parking space area obtains the actual coordinates of the four vertices through on-site calibration, and then converts them into image coordinates in the field of view captured by the camera through image processing technology. These vertices include the two front corners and two rear corners of the parking space, which together define the rectangular area of the parking space in the camera's field of view. This calibration helps the processor to perform subsequent image processing. During frequency analysis, the range of the parking space can be precisely located, ensuring that only the content within the designated area is analyzed when detecting wheels. For example, in a standard-sized parking space, technicians can use a rangefinder to mark the four vertices of the parking space, such as the upper left, upper right, lower left, and lower right corners. These coordinates on the screen are then recorded using image processing software for retrieval during parking detection. This coordinate information can be stored in the system's database and synchronized with the video frame information to ensure the real-time performance and accuracy of the detection. Preferably, during system initialization, multiple tests and adjustments should be performed to ensure that the camera's angle and height configuration perfectly cover the marked parking space area.
[0060] Furthermore, the pixel data of the video frame to be detected is input into the input layer of the wheel detection model. The wheel detection model performs layer-by-layer feature extraction and analysis to generate the position coordinates and confidence score of the bounding box of each vehicle wheel image.
[0061] As a preferred embodiment of the above, in a specific implementation, the video frame to be detected is first processed into pixel data, and then input into the input layer of the wheel detection model. The wheel detection model adopts a deep convolutional neural network structure to effectively process large-scale image data and extract multi-level features. The wheel detection model contains multiple layers of convolutional and pooling layers, which extract features from the input video frame data layer by layer. The initial layer is responsible for extracting low-level features, such as edges and basic shapes, while the intermediate and high-level layers focus on more specific structural features, enabling the model to more accurately identify the circular wheel structure. Through this hierarchical feature extraction, each layer can perform refined analysis of the input data. At the same time, the fully connected layer is combined to comprehensively process the output results of each layer to generate accurate positioning information. Preferably, the processed video frame pixel data is used for feature analysis in the model. Subsequently, the bounding box coordinates of the wheel image are returned. The bounding box coordinates include the position of the wheel in the video frame, which can effectively reflect the actual location of the wheel. The reliability of the recognition result is represented by a confidence score. These values are generated by combining the output of a multi-layer network with a predefined wheel pattern to ensure high accuracy and low error rate in the recognition process. For example, when processing outdoor parking lot videos, the bounding box coordinates of several possible wheels and their corresponding confidence scores can be obtained through model prediction. For each bounding box, when the confidence score exceeds a preset threshold, the bounding box is regarded as a valid wheel detection result and further used for parking status judgment. Through layer-by-layer feature extraction and analysis, accurate detection and localization of vehicle wheels are achieved, effectively supporting the stability of the entire parking status judgment process.
[0062] Furthermore, the wheel detection model includes:
[0063] Collect and construct a training sample set of video frames including vehicle wheel images;
[0064] For each frame of the training sample set, the precise location of all wheel targets is marked, and corresponding label data is generated;
[0065] The training sample set with labeled data is input into the wheel detection deep learning network architecture;
[0066] Based on the backpropagation algorithm and gradient descent optimizer, the weight parameters of the training wheel detection deep learning network architecture are iteratively adjusted until the loss function of the wheel detection model on the validation set converges to a preset range, thus obtaining the wheel detection model.
[0067] As a preferred embodiment of the above, firstly, video frames including vehicle wheel images are collected as a training sample set. Samples are collected using various types of camera equipment, different lighting conditions, and diverse angles to ensure coverage of various real-world scenarios, thereby improving the model's versatility and robustness. During sample labeling, the positions of all wheel targets are precisely marked in each frame, forming a label dataset containing detailed information such as location bounding boxes, shape, size, and specific wheel features. Next, the training sample set with labeled data is input into a wheel detection deep learning network architecture, typically constructed using a convolutional neural network. Then, training is performed using a backpropagation algorithm and a gradient descent optimizer. This step involves iteratively adjusting the network's weight parameters layer by layer. This iteration optimizes the parameters by calculating the error between the actual output and the expected output, enabling the model to more accurately distinguish vehicle wheels in images. Preferably, a validation set is used during training to monitor changes in the loss function. A preset loss range is set to determine the effectiveness of the training and whether the expected results have been achieved. The training process ends when the loss function converges to the preset range, ultimately obtaining an optimized wheel detection model that can quickly and accurately identify and locate wheels in images.
[0068] Furthermore, such as Figure 2 As shown, when a vehicle wheel image is recognized, the bounding box generated is a rectangle, and the position coordinates are defined by the center of the rectangle's upper left and lower right corners.
[0069] As a preferred embodiment of the above embodiments, during wheel detection, each identified vehicle wheel image is assigned a rectangular bounding box that accurately surrounds the detected wheel region to facilitate further analysis and application. The rectangular bounding box is automatically generated through feature map analysis in a deep learning model, and its range is defined using the top-left and bottom-right vertices. Preferably, the center position coordinates of the rectangular bounding box are calculated by extracting the coordinates of the two vertices, namely the top-left and bottom-right vertices. Calculating the center position facilitates a unified reference for wheel positions and simplifies the computational complexity of subsequent parking state determination. Specifically, the coordinates of the top-left vertex of the rectangular bounding box... The coordinates of the bottom right corner vertex represent the starting point of the wheel frame in the image, while the coordinates of the bottom right corner vertex provide the ending point of the frame. By using these vertex coordinates, the center point of the rectangle is calculated, reflecting the uniform and standardized position of the vehicle wheels in the image, which facilitates the subsequent analysis of parking space occupancy. For example, in a typical parking lot scenario, images of the wheels of a vehicle when parked are captured by a camera. After analysis by the wheel detection model, the system generates a rectangle for each wheel image and calculates its center position coordinates. In this way, the parking management system can use this center position data to accurately determine whether the vehicle is in the designated parking area and adjust the parking status notification accordingly.
[0070] Furthermore, such as Figure 3 As shown, the candidate landing point line is a straight line segment connecting any two different virtual ground contact points. In a physical sense, the straight line segment represents the projection line of the two wheels of a single vehicle on the ground. The orientation and length of the straight line segment characterize the vehicle's posture in the parking scenario.
[0071] As a preferred embodiment of the above, the key lies in utilizing the coordinates of the center point of the identified wheel image, i.e., the virtual ground contact point of the vehicle wheel, to generate candidate landing point lines by connecting these points. Each straight line segment connects two different virtual contact points, effectively reflecting the wheel projection line of the vehicle on the ground. During the detection process, for each identified vehicle, all possible combinations of virtual contact points are automatically calculated. This combination method can completely cover all candidate landing point lines, providing sufficient geometric information for further parking status analysis. Each candidate landing point line has a typical physical representation function; its direction indicates the vehicle's driving direction within the parking space, and its length corresponds to the vehicle's wheelbase. It provides detailed analysis of vehicle posture, and can quickly identify and respond to situations where the vehicle is significantly offset from the parking space or parked at an angle. Preferably, while generating these straight line segments, it utilizes other geometric data generated in the wheel recognition model and the camera's viewing angle parameters to optimize the realistic representation of the straight line segments. For example, it can accurately determine the vehicle's location based on this information under different lighting conditions during the day and night. For instance, in narrow city street parking lots, vehicles often need to complete parking operations within limited space. Through the above method, it can identify any violations of parking regulations, such as lateral parking and exceeding the parking space boundaries, and can promptly notify the driver and the parking management system to make adjustment decisions.
[0072] Furthermore, based on the length of each candidate landing point connection line, the length of the line segment that falls into the target parking space is compared and obtained.
[0073] Based on the length of the line segment that falls into the target parking space and the total length of the target parking space, calculate the corresponding occupancy ratio for each candidate landing point line.
[0074] As a preferred embodiment of the above, firstly, the geometric range of each target parking space is obtained through a pre-marked parking area, and all detected virtual ground contact points are combined to generate multiple candidate landing point lines. These lines physically represent the possible positions and directions of the vehicle. The key is to identify the overlapping parts of these lines with the target parking space. Through geometric calculation, the length of the intersection of each candidate landing point line with the parking space is accurately extracted. Next, the ratio of this intersection line segment to the total length of the target parking space is calculated to measure the vehicle occupancy represented by the line. Specifically, the occupancy ratio value refers to the result of dividing the intersection segment length by the total length of the parking space. This value can directly reflect whether the vehicle exceeds the parking space boundary or whether its parking angle is compliant. Preferably, the parking status is automatically determined by comparing the occupancy ratio value with a preset threshold. When the occupancy ratio value exceeds a certain threshold, it indicates that although the vehicle is in the parking space, it may not be parked completely compliantly, thereby triggering corresponding prompts and adjustment measures. Conversely, if the ratio value is good, the system confirms that the parking is reasonable and no intervention is required.
[0075] Furthermore, the percentage of occupancy includes:
[0076] The formula for calculating the occupancy ratio is:
[0077] IOU ij =(l ij ∩S) / l ij (i,j = 1, 2, ..., m);
[0078] Among them, l ij Let S represent the total length of the line connecting the i-th and j-th virtual ground contact points to form candidate landing points, and let S represent the area of the target parking space. ij ∩S represents the length of the line segment that falls within the geometric region S when connecting candidate landing points, IOU ij This is the calculated occupancy percentage.
[0079] As a preferred embodiment of the above, firstly, the geometric region of the target parking space is identified and its area is calculated. This region has been explicitly marked using the image coordinate system. During image detection, the virtual ground contact points obtained by identifying the vehicle wheels can form several candidate contact point lines. The total length of each candidate line can be obtained by calculating the distance connecting the two virtual contact points. Next, the length of the intersection line segment between the candidate contact point lines and the target parking space region needs to be calculated. Through geometric analysis, the overlapping part of each candidate line segment with the geometric region of the parking space is calculated. This part of the length is the actual occupied length within the target parking space region. In the calculation process, the boundaries of the parking space must be fully considered to ensure the accuracy of the line segment length. Then, the occupancy ratio value is obtained through the calculation formula of the occupancy ratio value, that is, the ratio of the overlap length of the candidate line segment in the target parking space to its total length. In this way, an intuitive ratio value can be obtained, which specifically reflects the fitting level of each vehicle's possible parking position relative to the parking area. Preferably, a threshold value of the occupancy ratio value is set so that potentially illegally parked vehicles can be easily filtered out in subsequent processes. For example, if the ratio value of a candidate line segment is lower than the threshold, the system will prompt an anomaly, which will constantly monitor whether the vehicle is parked in accordance with regulations.
[0080] Example 2;
[0081] Based on the same inventive concept as the roadside parking state determination method based on wheel detection in the foregoing embodiments, the present invention also provides a roadside parking state determination system based on wheel detection, the system comprising:
[0082] The video acquisition module acquires video frames to be detected that cover the target parking space. The geometric area of the target parking space is defined by a set of pre-calibrated position coordinates.
[0083] The position recognition module inputs the video frame to be detected into the wheel detection model, performs forward reasoning based on the wheel detection model, and outputs the position information of each identified vehicle wheel image.
[0084] The center determination module calculates and extracts the coordinates of the center point of the bottom edge of the bounding box of each vehicle wheel image, and determines the center point coordinates as the virtual ground contact point corresponding to the vehicle wheel.
[0085] The occupancy calculation module iterates through all virtual ground contact points, combines the virtual ground contact points in pairs to generate a set of candidate landing point lines representing potential vehicle locations, and calculates the occupancy ratio value for each candidate landing point line.
[0086] The status judgment module compares each occupancy ratio value with a preset parking status judgment threshold. If at least one occupancy ratio value is greater than the parking status judgment threshold, the target parking space is determined to be occupied; otherwise, the target parking space is determined to be unoccupied.
[0087] The adjustment system described above in this invention can effectively realize a roadside parking state judgment method based on wheel detection, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.
[0088] Furthermore, the location recognition module includes:
[0089] The image training unit collects and constructs a training sample set of video frames, including images of vehicle wheels.
[0090] The label correspondence unit marks the precise location of all wheel targets for each frame of the image in the training sample set and generates corresponding label data;
[0091] The architecture generation unit inputs the training sample set with labeled data into the wheel detection deep learning network architecture;
[0092] The model generation unit iteratively adjusts the weight parameters of the training wheel detection deep learning network architecture based on the backpropagation algorithm and gradient descent optimizer until the loss function of the wheel detection model on the validation set converges to a preset range, thus obtaining the wheel detection model.
[0093] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.
[0094] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.
Claims
1. A method for determining roadside parking status based on wheel detection, characterized in that, The method includes: Acquire the video frames to be detected that cover the target parking space, wherein the geometric area of the target parking space is defined by a set of pre-calibrated position coordinates; The video frame to be detected is input into the wheel detection model, and forward reasoning is performed based on the wheel detection model to output an image of each identified vehicle wheel. Calculate the center point coordinates of each vehicle wheel image, and determine the center point coordinates as the virtual ground contact point of the corresponding vehicle wheel; Traverse all the virtual ground contact points, combine the virtual ground contact points in pairs to generate a set of candidate landing point lines representing potential vehicle locations, and calculate the occupancy ratio value for each candidate landing point line. Each of the occupancy ratio values is compared with a preset parking status determination threshold. If at least one of the occupancy ratio values is greater than the parking status determination threshold, the target parking space is determined to be occupied; otherwise, the target parking space is determined to be unoccupied.
2. The method for determining roadside parking status based on wheel detection according to claim 1, characterized in that, The geometric region of the target parking space is defined by the coordinates of four vertices, which are labeled and recorded based on the image coordinate system to determine the precise coordinates of the four vertices of the target parking space.
3. The method for determining roadside parking status based on wheel detection according to claim 1, characterized in that, The pixel data of the video frame to be detected is input into the input layer of the wheel detection model. The wheel detection model performs layer-by-layer feature extraction and analysis to generate the position coordinates and confidence score of the bounding box of each vehicle wheel image.
4. The method for determining roadside parking status based on wheel detection according to claim 3, characterized in that, The wheel detection model includes: Collect and construct a training sample set of video frames including images of the vehicle wheels; For each frame of the training sample set, the precise location of all wheel targets is marked, and corresponding label data is generated; The training sample set with the labeled data is input into the wheel detection deep learning network architecture; The weight parameters of the deep learning network architecture for wheel detection are iteratively adjusted using the backpropagation algorithm and gradient descent optimizer until the loss function of the wheel detection model on the validation set converges to a preset range, thus obtaining the wheel detection model.
5. The method for determining roadside parking status based on wheel detection according to claim 3, characterized in that, When the vehicle wheel image is recognized, the bounding box generated is a rectangle, and the position coordinates are defined by the center of the coordinates of the upper left corner vertex and the lower right corner vertex of the rectangle.
6. The method for determining roadside parking status based on wheel detection according to claim 1, characterized in that, The candidate landing point line is a straight line segment connecting any two different virtual ground contact points. In a physical sense, the straight line segment represents the projection line of the two wheels of a single vehicle on the ground. The orientation and length of the straight line segment characterize the vehicle's posture in the parking scenario.
7. The method for determining roadside parking status based on wheel detection according to claim 1, characterized in that, Based on the length of each candidate landing point line, the length of the line segment that falls into the target parking space is obtained by comparison. Based on the length of the line segment falling into the target parking space and the total length of the target parking space, the corresponding occupancy ratio value is calculated for each of the candidate landing points.
8. The method for determining roadside parking status based on wheel detection according to claim 7, characterized in that, The occupancy ratio includes: The formula for calculating the occupancy ratio is: IOU ij =(l ij ∩S) / l ij (i,j=1,2……m)); Among them, l ij The length of the line connecting the i-th and j-th virtual ground contact points represents the total length of the candidate landing points, and the area of the target parking space represents the total length of the line connecting the i-th and j-th virtual ground contact points. ij ∩S represents the length of the line segment that falls within the geometric region S, where IOU ij The calculated occupancy ratio value.
9. A roadside parking status determination system based on wheel detection, characterized in that, The system includes: The video acquisition module acquires video frames to be detected that cover the target parking space. The geometric area of the target parking space is defined by a set of pre-calibrated position coordinates. The position recognition module inputs the video frame to be detected into the wheel detection model, performs forward reasoning based on the wheel detection model, and outputs the position information of each identified vehicle wheel image. The center determination module calculates and extracts the coordinates of the center point of the bottom edge of the bounding box of each vehicle wheel image, and determines the center point coordinates as the virtual ground contact point corresponding to the vehicle wheel. The occupancy calculation module iterates through all virtual ground contact points, combines the virtual ground contact points in pairs to generate a set of candidate landing point lines representing potential vehicle locations, and calculates the occupancy ratio value for each candidate landing point line. The status judgment module compares each occupancy ratio value with a preset parking status judgment threshold. If at least one occupancy ratio value is greater than the parking status judgment threshold, the target parking space is determined to be occupied; otherwise, the target parking space is determined to be unoccupied.
10. The roadside parking state determination system based on wheel detection according to claim 9, characterized in that, The location recognition module includes: The image training unit collects and constructs a training sample set of video frames, including images of vehicle wheels. The label correspondence unit marks the precise location of all wheel targets for each frame of the image in the training sample set and generates corresponding label data; The architecture generation unit inputs the training sample set with labeled data into the wheel detection deep learning network architecture; The model generation unit iteratively adjusts the weight parameters of the training wheel detection deep learning network architecture based on the backpropagation algorithm and gradient descent optimizer until the loss function of the wheel detection model on the validation set converges to a preset range, thus obtaining the wheel detection model.