Road parking specification degree evaluation method based on unmanned aerial vehicle image

Abstract

The invention provides a road parking specification degree evaluation method based on an unmanned aerial vehicle image. The method comprises the following steps: constructing a road parking specification degree evaluation platform; the road parking specification degree evaluation platform comprises a vehicle boundary detection network model, a vehicle key point detection network model and a parking space segmentation network model which are obtained through pre-construction and training; and according to the vehicle key point position projection coordinates and the angular point position coordinates of the parking space wireframe, vehicle parking specifications are evaluated. According to the invention, the problems of low efficiency, high labor cost and the like in traditional parking specification management are solved, the automation level and accuracy of parking specification management are improved, and an effective parking specification management means is provided for urban traffic management departments.

Description

Technical Field

[0001] The present invention belongs to the technical field of parking standardization evaluation, and in particular relates to a road parking standardization evaluation method based on drone images. Background Art

[0002] With the continuous growth of the number of motor vehicles, parking problems have become increasingly prominent, especially in densely populated and vehicle-heavy urban areas. Limited off-street parking facilities can no longer meet the demand. More and more roads have begun to divide parking spaces on both sides of the lanes to relieve parking pressure. However, this practice has led to the following new problems: 1. Traffic disorder: Irregular parking of vehicles disrupts traffic order, making it difficult for other vehicles to judge the use of parking spaces and road traffic conditions, increasing the risk of traffic disorder. Vehicles parked on the front and back lines or on the left and right lines may occupy adjacent parking spaces or lanes, making it impossible for other vehicles to park or pass normally. 2. Increase management difficulty and affect the image of the city: For traffic management departments, irregular parking behavior will increase workload and management difficulty, and make management departments invest more resources to supervise and manage parking behavior to ensure traffic order and public safety. At the same time, irregular parking behavior leaves the outside world with the impression of poor urban management, reducing the overall image and civilization of the city. For tourists and outsiders, irregular parking behavior affects their evaluation and impression of the city.

[0003] Therefore, how to simply and accurately detect and evaluate parking regulations to assist traffic management departments in managing road driving behavior is a key issue that needs to be addressed at present. Summary of the invention

[0004] In view of the defects of the prior art, the present invention provides a road parking standardization evaluation method based on drone images, which can effectively solve the above problems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] The present invention provides a method for evaluating road parking standardization based on drone images, comprising the following steps:

[0007] Step S1, constructing a road parking standardization evaluation platform; the road parking standardization evaluation platform includes a pre-constructed and trained vehicle boundary detection network model, a vehicle key point detection network model and a parking space segmentation network model;

[0008] Step S2, using a drone to inspect the road vehicle parking area, collecting vehicle parking area images with geographic coordinates and timestamps, and transmitting the collected vehicle parking area images to the road parking standardization evaluation platform in real time;

[0009] Step S3, the vehicle boundary detection network model of the road parking standardization evaluation platform performs vehicle boundary detection on the vehicle parking area image, and performs a minimum region of interest image acquisition operation for each detected vehicle boundary, the method is as follows:

[0010] For each detected vehicle boundary, the boundary is enlarged by a set multiple to obtain a minimum region of interest image including the complete vehicle and the complete parking space where the vehicle is located, and then the minimum region of interest image is cut out from the vehicle parking area image, and the minimum region of interest image is sent to the vehicle key point detection network model and the parking space segmentation network model respectively;

[0011] Step S4, establishing an image coordinate system in the minimum area of ​​interest image, the method is: taking the upper left corner point of the minimum area of ​​interest image as the coordinate origin O, taking the image width direction as the positive x direction, and taking the image length direction as the positive y direction, to establish the image coordinate system; based on the image coordinate system, the vehicle key point detection network model performs vehicle key point detection on the minimum area of ​​interest image, and extracts the left front light position coordinates (x1, y1), the left rear light position coordinates (x2, y2), the right rear light position coordinates (x3, y3) and the right front light position coordinates (x4, y4);

[0012] Step S5, the parking space segmentation network model performs parking space semantic segmentation on the minimum region of interest image, and fits the parking space wireframe, thereby obtaining the corner point position coordinates of the parking space wireframe, which are: left front corner point position coordinates (a1, b1), left rear corner point position coordinates (a2, b2), right rear corner point position coordinates (a3, b3) and right front corner point position coordinates (a4, b4);

[0013] Step S6, calibrating the image through a calibration plate to obtain a conversion matrix M between the coordinate system of the key point coordinates of the vehicle and the coordinate system of the parking space wireframe plane;

[0014] Using the transformation matrix M, the key point position coordinates of the vehicle extracted in step S4 are projected into the parking space wireframe plane to obtain the projection coordinates of the key point position of the vehicle on the parking space wireframe plane, which correspond to: (x'1, y'1), (x'2, y'2), (x'3, y'3) and (x'4, y'4);

[0015] Step S7, evaluating the parking specification of the vehicle according to the projection coordinates of the key point positions of the vehicle and the corner point position coordinates of the parking space wireframe in step S5.

[0016] Preferably, in step S3, each of the vehicle boundaries detected by the vehicle boundary detection network model is specifically a circumscribed rectangular frame of the vehicle.

[0017] Preferably, in step S6, the image is calibrated by a calibration board to obtain the conversion matrix M between the coordinate system where the vehicle key point coordinates are located and the coordinate system where the parking space line frame plane is located, specifically as follows:

[0018] The image is calibrated by a calibration board to obtain the coordinates of the four corner points of the calibration board (x01, y01), (x02, y02), (x03, y03), (x04, y04) on the plane of the vehicle key points and the coordinates of the four corner points of the calibration board (x11, y11), (x12, y12), (x13, y13), (x14, y14) in the parking space line frame plane;

[0019] Let vector x1 = (x11, x12, x13, x14), vector y1 = (y11, y12, y13, y14), vector x0 = (x01, x02, x03, x04), vector y0 = (y01, y02, y03, y04); According to the following formula:

[0020]

[0021] The conversion matrix M is obtained; where, m 11 , m 12 , m 13 , m 21 , m 22 , m 23 , m 31 and m 32 , are all conversion parameters of the conversion matrix.

[0022] Preferably, step S7 is specifically as follows:

[0023] Step S7.1, reverse parking judgment:

[0024] Calculate the average value m1 of the vertical coordinates of the two front points of the vehicle head as m1 = 1 / 2(y'1 + y'4) and the average value m2 of the vertical coordinates of the two rear points of the vehicle tail as m2 = 1 / 2(y'2 + y'3);

[0025] If m1 < m2, it is determined as forward parking; if m1 > m2, it is determined as reverse parking;

[0026] If it is reverse parking, then execute step S7.3; if it is non-reverse parking, execute step S7.2;

[0027] Step S7.2, judging parking on the line:

[0028] Determine whether the vehicle frame line composed of the four points (x'1, y'1), (x'2, y'2), (x'3, y'3) and (x'4, y'4) intersects with the parking space frame line composed of the four points (a1, b1)(a2, b2)(a3, b3)(a4, b4). If the vehicle frame line is completely within the parking space frame line and there is no intersection, it is concluded that the parking is not on the line, and step S7.3 is executed; otherwise, it is concluded that the parking is on the line, and step S7.3 is executed;

[0029] Step S7.3, parking standardization score:

[0030] Step S7.3.1, reverse parking scoring algorithm:

[0031] If it is reverse parking, the parking standardization score Score = 0, and the process ends;

[0032] Step S7.3.2, scoring algorithm for parking on the line:

[0033] If the parking is done on the line, the area S1 where the vehicle exceeds the parking space and the area S2 where the vehicle does not exceed the parking space are calculated;

[0034] Then the proportion of the car body over the line is IOU = S2 / (S1+S2);

[0035] The maximum score limit for vehicles parked on the line is 60. The formula Score = 60*IOU is used to get the parking standardization score Score, and the process ends;

[0036] Step S7.3.3, non-line-crossing parking scoring algorithm:

[0037] If it is non-line parking, the angle θ between the vehicle edge line and the parking space edge line is obtained according to the vehicle edge line of the vehicle frame line and the parking space edge line of the parking space frame line. The formula Score = 100-(40*θ) / 6 is used to obtain the parking standardization score Score, and the process ends.

[0038] The road parking standardization evaluation method based on drone images provided by the present invention has the following advantages:

[0039] This paper proposes a road parking standardization evaluation method based on drone images to achieve efficient and accurate roadside parking standardization evaluation. This method is based on high-definition image data collected by drones, combined with advanced parking line detection technology and vehicle posture detection technology, and can accurately detect and evaluate the parking standardization of vehicles. This innovative method is of great significance for optimizing urban traffic management and improving road use efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0041] Figure 1 A diagram of a road parking standardization evaluation system based on drone images provided in an embodiment of the present invention;

[0042] Figure 2 A flowchart of a road parking standardization evaluation system based on drone images provided by an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of perspective transformation of a road parking standardization evaluation system based on drone images provided by an embodiment of the present invention;

[0044] Figure 4 A model diagram of vehicle parking along an urban road according to a road parking standardization evaluation system based on drone images provided in an embodiment of the present invention;

[0045] Figure 5 A key point marking model diagram of a road parking standardization evaluation system based on drone images provided in an embodiment of the present invention;

[0046] Figure 6 A normal parking model diagram of a road parking standardization evaluation system based on drone images provided in an embodiment of the present invention;

[0047] Figure 7 An embodiment of the present invention provides a line parking model diagram of a road parking standardization evaluation system based on drone images. DETAILED DESCRIPTION

[0048] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] This invention aims to solve the shortcomings of current technologies and proposes an efficient and accurate roadside parking standard evaluation method. This method is based on high-definition image data collected by drones, combined with advanced parking line detection technology and vehicle posture detection technology, and can accurately detect and evaluate the parking standardization of vehicles. This innovative method is of great significance for optimizing urban traffic management and improving road use efficiency.

[0050] The road parking standardization evaluation method based on drone images provided by the present invention has the following advantages:

[0051] 1. By processing the images taken by drones in real time, the evaluation results of road parking standards can be quickly obtained, providing timely decision support for traffic management departments. In addition, the drone image evaluation method is not affected by human emotions and subjective judgments, and the evaluation results are more objective and fair. The same set of evaluation standards can be applied to parking scenes at different times and locations to ensure the fairness of the evaluation.

[0052] 2. The use of drone image evaluation methods can reduce the workload of manual inspections and reduce labor costs. At the same time, it can avoid safety hazards and traffic disruptions caused by manual inspections. The evaluation results can be visualized in the form of charts, maps, etc., which is convenient for traffic management departments to intuitively understand parking regulations.

[0053] Through data visualization, we can timely discover areas and time periods where parking problems are concentrated, providing data support for optimizing parking management.

[0054] See also Figure 1 The present invention provides a method for evaluating road parking standardization based on drone images, comprising the following steps:

[0055] Step S1, constructing a road parking standardization evaluation platform; the road parking standardization evaluation platform includes a pre-constructed and trained vehicle boundary detection network model, a vehicle key point detection network model and a parking space segmentation network model;

[0056] Step S2, using a drone to inspect the road vehicle parking area, collecting vehicle parking area images with geographic coordinates and timestamps, and transmitting the collected vehicle parking area images to the road parking standardization evaluation platform in real time;

[0057] Step S3, the vehicle boundary detection network model of the road parking standardization evaluation platform can use the MMDet neural network to perform vehicle boundary detection on the vehicle parking area image, and detect each vehicle boundary, specifically the circumscribed rectangular frame of the vehicle; for each detected vehicle boundary, a minimum region of interest image acquisition operation is performed, and the method is:

[0058] For each detected vehicle boundary, it is enlarged by a set multiple to obtain a minimum area of ​​interest image that includes the complete vehicle and the complete parking space where it is located, and then the minimum area of ​​interest image is cropped from the vehicle parking area image, and the minimum area of ​​interest image is sent to the vehicle key point detection network model and the parking space segmentation network model respectively. The minimum area of ​​interest image is the subsequent processing target.

[0059] In practical applications, a training data set is established and annotated by using drones to take roadside parking images during road inspections, and the MMDet neural network is trained. The vehicle boundary is predicted based on the trained MMDet neural network to obtain the vehicle boundary coordinates.

[0060] Step S4, such as Figure 4 As shown, an image coordinate system is established in the minimum area of ​​interest image, and the method is: take the upper left corner point of the minimum area of ​​interest image as the coordinate origin O, take the image width direction as the positive x direction, and take the image length direction as the positive y direction to establish the image coordinate system; based on the image coordinate system, the vehicle key point detection network model performs vehicle key point detection on the minimum area of ​​interest image, and extracts the left front light position coordinates (x1, y1), the left rear light position coordinates (x2, y2), the right rear light position coordinates (x3, y3) and the right front light position coordinates (x4, y4); wherein, the vehicle key point detection network model can adopt the vehicle key point extraction network RTMPose model. Specifically, a training data set is established and annotated by using drones to patrol roads and shoot roadside parking images, and the RTMPose network model is trained; the obtained RTMPose network model is used to extract vehicle key points from the minimum area of ​​interest image.

[0061] Step S5, the parking space segmentation network model uses semantic segmentation technology to perform parking space semantic segmentation on the minimum region of interest image, and fits the parking space wireframe through polygons, thereby obtaining the corner point position coordinates of the parking space wireframe, which are: left front corner point position coordinates (a1, b1), left rear corner point position coordinates (a2, b2), right rear corner point position coordinates (a3, b3) and right front corner point position coordinates (a4, b4);

[0062] Specifically, a training data set is established and annotated by using drones to patrol roads and take roadside parking images, and a parking space segmentation network model is trained; the parking space segmentation network model adopts Mask-RCNN; the trained Mask-RCNN model is used to extract the corner point positions of the parking space wireframe; specifically, the mask segmentation of the target parking space is realized through a neural network to obtain the target parking space border and target category.

[0063] Use the obtained network model to predict the key points (x2, y2) (x3, y3) (x4, y4) of the region of interest image in step 1 to obtain the key point coordinate information of the vehicle

[0064] Step S6, calibrating the image through a calibration plate to obtain a conversion matrix M between the coordinate system of the key point coordinates of the vehicle and the coordinate system of the parking space wireframe plane;

[0065] Using the transformation matrix M, the key point position coordinates of the vehicle extracted in step S4 are projected into the parking space wireframe plane to obtain the projection coordinates of the key point position of the vehicle on the parking space wireframe plane, which correspond to: (x'1, y'1), (x'2, y'2), (x'3, y'3) and (x'4, y'4); Figure 5 As shown, it is a model diagram of vehicle key point marking;

[0066] In this step, the image is calibrated by the calibration plate to obtain the transformation matrix M between the coordinate system of the vehicle key point coordinates and the coordinate system of the parking space wireframe plane. Figure 3 , specifically:

[0067] An upper calibration plate is established on the plane of the vehicle key points; a lower calibration plate is established on the plane of the parking space wireframe; the upper calibration plate and the lower calibration plate are parallel and completely aligned;

[0068] The image is calibrated by the upper calibration plate to obtain the coordinates of the four corner points of the upper calibration plate located in the plane of the vehicle key points (x01, y01)(x02, y02)(x03, y03)(x04, y04); the image is calibrated by the lower calibration plate to obtain the coordinates of the four corner points of the lower calibration plate located in the plane of the parking space wireframe (x11, y11)(x12, y12)(x13, y13)(x14, y14);

[0069] Let vector x1 = (x11, x12, x13, x14), vector y1 = (y 11, y 12, y 13, y 14), vector x0 = (x01, x02, x03, x04), vector y0 = (y01, y02, y03, y04); according to the following formula:

[0070]

[0071] Get the transformation matrix M; where m 11 ,m 12 ,m 13 ,m 21 ,m 22 ,m 23 ,m 31 and m 32 , are the transformation parameters of the transformation matrix.

[0072] Step S7, evaluating the parking specification of the vehicle according to the projection coordinates of the key point positions of the vehicle and the corner point position coordinates of the parking space wireframe in step S5.

[0073] Combination Figure 2 , step S7 specifically comprises:

[0074] Step S7.1, reverse parking determination:

[0075] Calculate the average value m1 of the vertical coordinates of the two points at the front of the vehicle, m1 = 1 / 2(y’1 + y’4), and the average value m2 of the vertical coordinates of the two points at the rear of the vehicle, m2 = 1 / 2(y’2 + y’3);

[0076] If m1 < m2, it is determined as forward parking; if m1 > m2, it is determined as reverse parking;

[0077] If it is reverse parking, execute step S7.3; if it is non-reverse parking, execute step S7.2;

[0078] Step S7.2, judging parking over the line:

[0079] Judge whether the vehicle frame line formed by the four points (x’1, y’1), (x’2, y’2), (x’3, y’3) and (x’4, y’4) intersects with the parking space line frame formed by the four points (a1, b1), (a2, b2), (a3, b3) and (a4, b4). If the vehicle frame line is completely inside the parking space line frame and there is no intersection, it is concluded that the parking is not over the line. As Figure 6 shown, it is a schematic diagram of non-over-the-line parking, and execute step S7.3; otherwise, as Figure 7 shown, it is concluded that the parking is over the line, and execute step S7.3;

[0080] Step S7.3, parking standard score Score:

[0081] Step S7.3.1, reverse parking scoring algorithm:

[0082] If it is reverse parking, the parking standard score Score = 0, and the process ends;

[0083] Step S7.3.2, over-the-line parking scoring algorithm:

[0084] If it is over-the-line parking, calculate the area S1 where the vehicle exceeds the parking space and the area S2 where the vehicle does not exceed the parking space; specifically, S1 is the total area of the part of the vehicle frame line outside the parking space frame, and S2 is the total area of the part of the vehicle frame line inside the parking space frame;

[0085] Then the proportion IOU of the part of the vehicle body exceeding the line = S2 / (S1 + S2);

[0086] The maximum score for a vehicle parked over the line is 60. Using the formula Score = 60 * IOU, the parking standard score Score is obtained, and the process ends;

[0087] Step S7.3.3, non-over-the-line parking scoring algorithm:

[0088] If it is non-line parking, the angle θ between the vehicle edge line and the parking space edge line is obtained according to the vehicle edge line of the vehicle frame line and the parking space edge line of the parking space frame line. The formula Score = 100-(40*θ) / 6 is used to obtain the parking standardization score Score, and the process ends.

[0089] In this step, the maximum value of the angle θ is:

[0090]

[0091] Therefore, in the present invention, for reverse parking, the standardization score is zero, for parking on the line, the final score of the standardization evaluation score is: score = 60*IOU, and for parking without hitting the line, the final standardization evaluation score is: score = 100-(40*θ) / 6.

[0092] Current technology cannot provide a parking standardization detection method with low management manpower cost, less installation equipment, and low construction and operation and maintenance costs, and there is no way to reflect the level of parking standardization in the form of a percentage score. The present invention aims to address the deficiencies in current technology and proposes an efficient and accurate roadside parking standardization evaluation method. The method is based on high-definition image data collected by drones, combined with advanced parking line detection technology and vehicle posture detection technology, and can accurately detect and evaluate the parking standardization of vehicles. This innovative method is of great significance for optimizing urban traffic management and improving road use efficiency.

[0093] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0094] 1. By processing the images taken by drones in real time, the evaluation results of road parking standards can be quickly obtained, providing timely decision support for traffic management departments. In addition, the drone image evaluation method is not affected by human emotions and subjective judgments, and the evaluation results are more objective and fair. The same set of evaluation standards can be applied to parking scenes at different times and locations to ensure the fairness of the evaluation.

[0095] 2. The use of drone image evaluation methods can reduce the workload of manual inspections and reduce labor costs. At the same time, it can avoid safety hazards and traffic disruptions caused by manual inspections. The evaluation results can be visualized in the form of charts, maps, etc., which is convenient for traffic management departments to intuitively understand parking regulations.

[0096] Through data visualization, we can timely discover areas and time periods where parking problems are concentrated, providing data support for optimizing parking management.

[0097] An embodiment is described below:

[0098] like Figure 1 As shown, the implementation of the present invention specifically requires the completion of three parts, namely model preparation, vehicle parking situation acquisition and parking scoring.

[0099] Step S1: Model preparation requires the use of parking images collected by drone inspections to create a training set, and the annotation of vehicle key points and parking space line masks needs to be completed. After completing the above work, the training set is used to train the vehicle boundary detection network, vehicle key point detection network, and parking space detection network. During the training process, the training parameters need to be appropriately adjusted to obtain the optimal network model.

[0100] Step S2: When obtaining the parking situation of the vehicle, the following steps need to be completed:

[0101] Step S2.1: Predict the vehicle boundary of the drone image according to the trained MMDet detector model, and obtain the diagonal coordinates (A1, B1) (A2, B2) of the rectangular box containing the vehicle;

[0102] Step S2.2: appropriately enlarge the rectangular frame, and crop the original image with the enlarged rectangular frame to obtain a cropped image of the minimum area of ​​interest that completely includes the vehicle and the parking space;

[0103] Step S2.3: Use the obtained RTMPose network model to predict the key points of the vehicle in the minimum area of ​​interest image, and obtain the key point coordinate information of the vehicle, which are: the left front light position coordinate (x1, y1), the left rear light position coordinate (x2, y2), the right rear light position coordinate (x3, y3) and the right front light position coordinate (x4, y4);

[0104] Step S2.4: Use the trained parking space segmentation network Mask-RCNN model to perform parking space target detection and image segmentation on the minimum region of interest image to obtain a semantic segmentation image of the parking space;

[0105] Step S2.5: The semantic segmentation image of the parking space is subjected to polygon fitting to obtain the corner point coordinate information of the parking space line, which are: the left front corner point position coordinate (a1, b1), the left rear corner point position coordinate (a2, b2), the right rear corner point position coordinate (a3, b3) and the right front corner point position coordinate (a4, b4);

[0106] Step S2.6: Calibrate the image through the calibration plate to obtain the coordinates of the four corner points of the calibration plate located on the vehicle key point plane (x01, y01) (x02, y02) (x03, y03) (x04, y04), and the coordinates of the four corner points of the calibration plate in the parking space plane (x11, y11) (x12, y12) (x13, y13) (x14, y14);

[0107] Substitute the above coordinates into the formula

[0108] Obtain the transformation matrix M of the coordinate system where the vehicle key point coordinates are located and the coordinate system where the parking space line plane is located;

[0109] Step S2.7: Project the vehicle key point coordinates obtained in step S2.3 into the parking space line plane through the transformation matrix M to obtain the projected coordinates of the vehicle key point positions in the parking space line frame plane, which are respectively: (x'1, y'1), (x'2, y'2), (x'3, y'3) and (x'4, y'4);

[0110] Step S2.8: Calculate the evaluation degree of the projected vehicle key point coordinate information and the parking space line coordinate information obtained in step S2.5.

[0111] Step 3: The specific judgment basis is as follows:

[0112] By calculation, the average values of the vertical coordinates of the two points at the front of the vehicle and the average values of the vertical coordinates of the two points at the rear of the vehicle can be obtained as m1 = 1 / 2(y'1 + y'4) and m2 = 1 / 2(y'2 + y'3);

[0113] If m1 > m2, it is determined as reverse parking, and the evaluation score is zero;

[0114] If m1 < m2, it is determined as forward parking, and it is necessary to further judge the parking standard degree and the angle θ between the vehicle side line and the parking space line.

[0115] Judge whether the vehicle is completely within the parking space line by the vehicle rectangle formed by the four points (x'1, y'1), (x'2, y'2), (x'3, y'3) and (x'4, y'4) and the parking space line rectangle formed by the four points (a1, b1)(a2, b2)(a3, b3)(a4, b4);

[0116] If it is parked with the wheels on the line, that is, there is an intersection between the parking space line frame line and the vehicle frame line, calculate the area S1 where the vehicle exceeds the parking space and the area S2 where the vehicle does not exceed the parking space; further calculate the proportion IOU = S2 / (S1 + S2) of the part where the vehicle body exceeds the line

[0117] For vehicles parked with the wheels on the line, the maximum score limit is 60, and the formula Score = 60 * IOU is used to calculate the final score;

[0118] If it is not parked with the wheels on the line, that is, the vehicle frame line is completely within the parking space line frame line, that is, calculate the angle θ between the vehicle side line constructed by using the key points and the parking space side line, and calculate the evaluation score by using the angle θ between the vehicle side line and the parking space side line: Score = 100 - (40 * θ) / 6

[0119] The present invention proposes a road parking standardization evaluation method based on drone images. The method uses a drone to inspect the road parking area along a preset route, and uses a deep learning algorithm to process the collected image data to realize automatic evaluation of the vehicle parking standardization. First, the vehicle boundary in the image is detected by the MMDet neural network, the region of interest is extracted, and then the position information of the vehicle key points is extracted by the RTMPose model. At the same time, the parking space line is fitted and the corner point position is obtained by semantic segmentation technology. Finally, the evaluation score of the vehicle parking standardization is calculated by combining the vehicle key point and parking space line information.

[0120] The present invention solves the problems of low efficiency and high labor cost in traditional parking standardization management, improves the automation level and accuracy of parking standardization management, and provides an effective parking standardization management method for urban traffic management departments.

[0121] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A road parking standardization evaluation method based on drone images, characterized in that: The following steps are involved: Step S1, constructing a road parking standardization evaluation platform; the road parking standardization evaluation platform includes a pre-constructed and trained vehicle boundary detection network model, a vehicle key point detection network model and a parking space segmentation network model; Step S2, using a drone to inspect the road vehicle parking area, collecting vehicle parking area images with geographic coordinates and timestamps, and transmitting the collected vehicle parking area images to the road parking standardization evaluation platform in real time; Step S3, the vehicle boundary detection network model of the road parking standardization evaluation platform performs vehicle boundary detection on the vehicle parking area image, and performs a minimum region of interest image acquisition operation for each detected vehicle boundary, the method is as follows: For each detected vehicle boundary, the boundary is enlarged by a set multiple to obtain a minimum region of interest image including the complete vehicle and the complete parking space where the vehicle is located, and then the minimum region of interest image is cut out from the vehicle parking area image, and the minimum region of interest image is sent to the vehicle key point detection network model and the parking space segmentation network model respectively; Step S4, establishing an image coordinate system in the minimum area of ​​interest image, the method is: taking the upper left corner point of the minimum area of ​​interest image as the coordinate origin O, taking the image width direction as the positive x direction, and taking the image length direction as the positive y direction, to establish the image coordinate system; based on the image coordinate system, the vehicle key point detection network model performs vehicle key point detection on the minimum area of ​​interest image, and extracts the left front light position coordinates (x1, y1), the left rear light position coordinates (x2, y2), the right rear light position coordinates (x3, y3) and the right front light position coordinates (x4, y4); Step S5, the parking space segmentation network model performs parking space semantic segmentation on the minimum region of interest image, and fits the parking space wireframe, thereby obtaining the corner point position coordinates of the parking space wireframe, which are: left front corner point position coordinates (a1, b1), left rear corner point position coordinates (a2, b2), right rear corner point position coordinates (a3, b3) and right front corner point position coordinates (a4, b4); Step S6, calibrating the image through a calibration plate to obtain a conversion matrix M between the coordinate system of the key point coordinates of the vehicle and the coordinate system of the parking space wireframe plane; Using the transformation matrix M, the key point position coordinates of the vehicle extracted in step S4 are projected into the parking space wireframe plane to obtain the projection coordinates of the key point position of the vehicle on the parking space wireframe plane, which correspond to: (x'1, y'1), (x'2, y'2), (x'3, y'3) and (x'4, y'4); Step S7, evaluating the parking specification of the vehicle according to the projection coordinates of the key point positions of the vehicle and the corner point position coordinates of the parking space wireframe in step S5.

2. The method for evaluating road parking standardization based on drone images according to claim 1 is characterized in that: In step S3, each of the vehicle boundaries detected by the vehicle boundary detection network model is specifically a circumscribed rectangular frame of the vehicle.

3. The method for evaluating road parking standardization based on drone images according to claim 1, characterized in that: In step S6, the image is calibrated by a calibration board to obtain the transformation matrix M between the coordinate system where the vehicle key point coordinates are located and the coordinate system where the parking space line frame plane is located, specifically as follows: The image is calibrated by a calibration board to obtain the coordinates of the four corner points of the calibration board on the vehicle key point plane (x01, y01), (x02, y02), (x03, y03), (x04, y04) and the coordinates of the four corner points of the calibration board within the parking space line frame plane (x11, y11), (x12, y12), (x13, y13), (x14, y14); Let vector x1 = (x11, x12, x13, x14), vector y1 = (y11, y12, y13, y14), vector x0 = (x01, x02, x03, x04), vector y0 = (y01, y02, y03, y04); According to the following formula: Get the transformation matrix M; where m 11 ,m 12 ,m 13 ,m 21 ,m 22 ,m 23 ,m 31 and m 32 , are the transformation parameters of the transformation matrix.

4. The method for evaluating road parking standardization based on drone images according to claim 1, characterized in that: Step S7 is specifically as follows: Step S7.1, reverse parking judgment: Calculate the average value m1 of the vertical coordinates of the two front points of the vehicle head as m1 = 1 / 2(y’1 + y’4) and the average value m2 of the vertical coordinates of the two rear points of the vehicle tail as m2 = 1 / 2(y’2 + y’3); If m1 < m2, it is determined as forward parking. If m1 > m2, it is determined as reverse parking; If it is reverse parking, then execute step S7.3; If it is not reverse parking, execute step S7.2; Step S7.2, on-line parking judgment: Judge whether there is an intersection between the vehicle frame line formed by the four points (x’1, y’1), (x’2, y’2), (x’3, y’3) and (x’4, y’4) and the parking space line frame line formed by the four points (a1, b1), (a2, b2), (a3, b3) and (a4, b4). If the vehicle frame line is completely within the parking space line frame line and there is no intersection, then draw the conclusion of non-on-line parking and execute step S7.3; otherwise, draw the conclusion of on-line parking and execute step S7.3; Step S7.3, parking standard score Score: Step S7.3.1, reverse parking scoring algorithm: If it is reverse parking, then the parking standard score Score = 0, and the process ends; Step S7.3.2, on-line parking scoring algorithm: If it is on-line parking, then calculate the area S1 where the vehicle exceeds the parking space and the area S2 where the vehicle does not exceed the parking space; Then the proportion IOU of the part of the vehicle body exceeding the line = S2 / (S1 + S2); The maximum score for on-line parked vehicles is limited to 60. Using the formula Score = 60 * IOU, the parking standard score Score is obtained, and the process ends; Step S7.3.3, non-on-line parking scoring algorithm: If it is non-on-line parking, then according to the vehicle side line of the vehicle frame line and the parking space side line of the parking space line frame line, the included angle θ between the vehicle side line and the parking space side line is obtained. Using the formula Score = 100 - (40 * θ) / 6, the parking standard score Score is obtained, and the process ends.

Images