Microwave radar auxiliary-based parked vehicle recognition method and storage medium

Abstract

The application discloses a parking vehicle identification method based on microwave radar assistance and a storage medium, and the method comprises the following steps: collecting radar data of a to-be-detected target through a microwave radar sensor, and performing preliminary category identification on the to-be-detected target according to the radar data to obtain a preliminary identification result; if the preliminary identification result is a car, an image of the to-be-detected target is captured through a camera device, and the image of the to-be-detected target is identified through a pre-set trained deep learning model to obtain the category of the to-be-detected target. The application can effectively reduce the invalid snapshot probability of the camera on the parking space.

Description

Technical Field

[0001] This invention relates to the field of vehicle identification technology, and in particular to a method and storage medium for identifying parked vehicles based on microwave radar assistance. Background Technology

[0002] Traditional vehicle detection relies on single environmental sensing devices (millimeter-wave radar, lidar, cameras, etc.), which are susceptible to interference from environmental factors, leading to poor vehicle detection and recognition performance. To address these issues, existing vehicle detection technologies often integrate various sensors based on the principle of complementary advantages, leveraging the applicability of different sensors in various complex environments. Camera sensors, used to acquire visible light images, offer high accuracy in perceiving the shape and category of objects, and camera-based sensing and image processing technologies are relatively mature. However, this technology is significantly affected by weather and environmental conditions in real-world applications. Millimeter-wave radar, on the other hand, features narrow antenna beamwidth, high resolution, wide bandwidth, and strong anti-interference capabilities. It is less affected by weather and environmental conditions and can accurately measure the speed and distance of objects. Its lower cost and ease of mass production have made vehicle detection technologies combining millimeter-wave radar and cameras increasingly popular. Compared to single-sensor detection technologies, this approach significantly improves vehicle detection performance while also posing greater challenges in areas such as deep information fusion processing of multi-source data and the functional coupling of sensor hardware.

[0003] Existing technologies have proposed vehicle detection techniques based on the fusion of millimeter-wave radar and machine vision. The processing mechanism of these techniques typically involves first obtaining a hypothetical target area based on millimeter-wave radar image detection results, then combining machine vision technology to perform image detection on the hypothetical target area, and finally using deep learning technology based on big data to identify vehicles from the image. This processing mechanism reduces the false detection rate of radar detection and the computational load of image visual detection to some extent. However, this vehicle recognition technology relies entirely on machine vision technology and cannot solve the problem of invalid vehicle captures by cameras, resulting in some invalid captured images occupying device storage space. Furthermore, this technique is prone to problems such as loss of effective information due to human factors and insufficient information mining during the data processing of various initial data sources, failing to achieve in-depth mining of effective information from various data sources. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a parking vehicle identification method and storage medium based on microwave radar assistance, which can effectively reduce the probability of invalid capture by cameras in parking spaces and improve the identification accuracy of parking vehicles.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a parking vehicle identification method based on microwave radar assistance, comprising:

[0006] Radar data of the target under test is acquired by a microwave radar sensor, and the target under test is initially classified based on the radar data to obtain a preliminary identification result.

[0007] If the preliminary identification result is a car, then an image of the target to be tested is captured by a camera device, and the image of the target to be tested is identified by a pre-trained deep learning model to obtain the category of the target to be tested.

[0008] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0009] The beneficial effects of this invention are as follows: by collecting radar data of the target to be tested, preliminary category identification is performed based on the radar data, and the capture identification is triggered only when the preliminary identification result is a car. This can effectively avoid problems such as capturing empty cars, capturing passing vehicles, and mistakenly capturing non-small and medium-sized cars due to interference from equipment and environmental factors, reduce the probability of invalid capture, enhance the anti-interference ability against multiple environmental variables, and realize intelligent identification of small and medium-sized cars with parking intentions. Attached Figure Description

[0010] Figure 1 This is a flowchart of a parking vehicle identification method based on microwave radar assistance according to the present invention;

[0011] Figure 2 This is a flowchart of the first part of Embodiment 1 of the present invention;

[0012] Figure 3 This is a schematic diagram showing the variation curves of echo signal amplitude between vehicle-type targets and non-vehicle-type targets under different radar detection frequencies.

[0013] Figure 4 This is a schematic diagram of the network structure of the YOLOv5 model in Embodiment 1 of the present invention;

[0014] Figure 5 This is a flowchart of the second part of Embodiment 1 of the present invention. Detailed Implementation

[0015] To explain the technical content, objectives, and effects of the present invention in detail, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0016] Please see Figure 1 A method for identifying parked vehicles based on microwave radar assistance, comprising:

[0017] Radar data of the target under test is acquired by a microwave radar sensor, and the target under test is initially classified based on the radar data to obtain a preliminary identification result.

[0018] If the preliminary identification result is a car, then an image of the target to be tested is captured by a camera device, and the image of the target to be tested is identified by a pre-trained deep learning model to obtain the category of the target to be tested.

[0019] As can be seen from the above description, the beneficial effects of the present invention are: it can effectively reduce the probability of invalid capture by the camera in the parking space and achieve effective capture of vehicles in the parking space.

[0020] Further, before acquiring radar data of the target under test through a microwave radar sensor and performing preliminary category identification of the target under test based on the radar data to obtain a preliminary identification result, the process further includes:

[0021] The system acquires attribute information for various types of moving objects and, based on the same test environment and movement path, collects radar data for each type of moving object at different detection frequencies, speed change states, and distances using a microwave radar sensor. The types of moving objects include cars, electric vehicles, and people. The attribute information includes length, width, and height. The speed change states include uniform movement and deceleration to a stop. The distance is the distance between the moving object and the transmission center of the microwave radar sensor. The radar data includes Doppler frequency shift, polarization state, and echo signal amplitude.

[0022] Based on the same radar data corresponding to different detection frequencies, the same speed change state, and the same distance for each type of moving object, curves of automobiles and other types of moving objects corresponding to the same radar data are plotted in a preset coordinate system. The horizontal axis of the coordinate system is the detection frequency, and the vertical axis is the same radar data.

[0023] Based on the curves of the same radar data corresponding to the car and other types of moving objects, determine whether the same radar data is valid;

[0024] If valid, the same radar data will be used as sample radar data.

[0025] A standard training sample set is generated based on the attribute information of each type of moving object and the corresponding sample radar data of different speed changes and different distances.

[0026] The preset machine learning model is trained based on the standard training sample set to obtain a trained machine learning model, which is a support vector machine model.

[0027] As described above, a single-factor analysis is first performed on each radar data to determine the effective radar data that can distinguish between car-type objects and non-car-type objects to a certain extent. Then, a linkage identification analysis is performed on each effective radar data to achieve preliminary category identification based on microwave radar technology.

[0028] Furthermore, the step of determining whether the same radar data is valid based on the curves of the vehicle and other types of moving objects corresponding to the same radar data specifically involves:

[0029] If a discrimination threshold curve exists corresponding to the same radar data, then the same radar data is determined to be valid. The portion of the discrimination threshold curve above the preset first threshold is located between the curves corresponding to the same radar data for the vehicle and other types of moving objects, and the discrimination threshold curve has an effective detection range. The recognition accuracy of moving object type identification based on the threshold point corresponding to the effective detection range on the discrimination threshold curve reaches the preset second threshold.

[0030] Furthermore, the step of using the same radar data as sample radar data specifically means:

[0031] Based on the morphological and extreme value characteristics of the curves corresponding to the same radar data for the vehicle and other types of moving objects, the optimal detection frequency corresponding to the same radar data is determined.

[0032] Based on the optimal detection frequency corresponding to the same radar data, the same radar data corresponding to different speed changes and different distances for each type of moving object is collected as sample radar data.

[0033] As described above, by determining the optimal detection frequency for effective radar data and then re-collecting radar data at that optimal detection frequency as sample data for subsequent model training, the accuracy of model recognition can be improved.

[0034] Furthermore, the preliminary category identification of the target under test based on the radar data to obtain a preliminary identification result specifically includes:

[0035] Based on the radar data of the target to be tested, the trained machine learning model is used to perform preliminary category identification of the target to be tested, and a preliminary identification result is obtained.

[0036] Furthermore, before identifying the image of the target to be tested using a pre-trained deep learning model to obtain the category of the target, the process further includes:

[0037] The captured images are obtained by using a camera device, and the captured images are cleaned to obtain a sample dataset;

[0038] Based on the preset training image size, data augmentation and adaptive anchor box calculation are performed on the sample dataset, and the images in the sample dataset are scaled, filled, labeled and format converted to obtain a standard image sample set.

[0039] Based on the standard image sample set and preset training hyperparameters, a preset deep learning model is trained to obtain a trained deep learning model.

[0040] Furthermore, the deep learning model is a YOLOv5 model, which includes an input terminal, a backbone network, a neck network, and an output terminal.

[0041] The input terminal is used to input images;

[0042] The backbone network is used to extract image features;

[0043] The neck network is used to fuse image features;

[0044] The output is used to predict the location and category of the target.

[0045] Furthermore, the loss function of the YOLOv5 model is the CIOU_Loss function.

[0046] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0047] Example 1

[0048] Please refer to Figure 2-5 Embodiment 1 of the present invention is: a parking vehicle identification method based on microwave radar assistance, which can be applied to urban smart parking projects and can effectively identify and capture evidence of small and medium-sized cars parked in parking spaces in a timely manner.

[0049] This method mainly consists of two parts. The first part involves acquiring radar data of the target using a microwave radar sensor and performing preliminary category identification based on the radar data to obtain a preliminary identification result. The second part involves capturing an image of the target using a camera device when the preliminary identification result indicates it is a vehicle. A pre-trained deep learning model is then used to identify the target image and determine its category. In essence, the preliminary identification result obtained from the radar data serves as a prerequisite for triggering the camera device to capture an image, reducing the probability of invalid captures. The captured image is then further processed for detection and identification to improve the accuracy of the identification.

[0050] like Figure 2 As shown, the first part specifically includes the following steps:

[0051] S101: Acquire attribute information for various types of moving objects, and based on the same test environment and movement path, collect radar data for each type of moving object at different detection frequencies, speed change states, and distances using a microwave radar sensor. The types of moving objects include cars (small and medium-sized cars), electric vehicles, and people. Attribute information includes length, width, and height. Speed ​​change states include (low-speed) uniform movement and deceleration to a stop. The distance is the distance between the moving object and the transmission center of the microwave radar sensor. Radar data includes Doppler shift, polarization, and echo amplitude.

[0052] Specifically, microwave radar data is generally represented as a two-dimensional point cloud, containing information such as (x, y) coordinates, RCS (radar cross section), and Doppler (object velocity). Traditional radar sensors mostly emphasize the detection capability of dynamic obstacles, acquiring information such as the speed, distance, and orientation of obstacles within the detectable range, but have not explored the feasibility of obstacle category identification.

[0053] To fully utilize the available identification information of microwave radar, this embodiment employs supervised classification. Based on the same test environment and movement path, microwave radar sensors are used to collect relevant characteristic information of different types of moving objects (pedestrians, small and medium-sized cars, electric vehicles, etc.) and their echoes at different distances from the radar transmission center, including Doppler shift, polarization, echo amplitude, echo phase, echo time, and radar angles.

[0054] The relevant characteristic information includes geometric characteristics, surface reflection characteristics, and behavioral state characteristics.

[0055] Geometric characteristics refer to the spatial geometry of a moving object, primarily manifested as differences in its length, width, and height. In this embodiment, the specific differences in the test object (based on the device's perspective) are as follows:

[0056] a) In terms of height: People ≥ Small and medium-sized cars > Electric vehicles;

[0057] b) In terms of width: Small and medium-sized cars > Electric vehicles ≥ People;

[0058] c) In terms of length: small and medium-sized cars > electric vehicles > people.

[0059] The behavioral state characteristics are manifested as changes in the velocity state of a moving object. In this embodiment, from the moment the moving object enters the effective detection range of the radar until it comes to a stop in the parking space, the test records two behavioral state processes: uniform motion at low speed and gradual deceleration to a standstill.

[0060] The surface reflection characteristics are mainly manifested in the differences in the degree of radar response (i.e. the energy intensity of the received radar signal) among different types of moving objects.

[0061] In addition, the raw data acquired by the radar is explained as follows:

[0062] (a) Doppler Shift: This is the frequency change caused by the target's motion, which can be used to understand the target's motion state;

[0063] (b) Polarization: This is the direction of the electric field vector in microwave transmission. The material properties of the target can be identified by the signals received under different polarization modes.

[0064] (c) Echo Amplitude: This is the intensity of the microwave signal received by the radar, which can be used to understand the target's reflectivity;

[0065] (d) Echo Phase: This is the phase of the microwave signal, which can be used to calculate the distance between the target and the radar;

[0066] (e) Echo Time: This is the time interval between the transmission and reception of the microwave signal, which can be used to calculate the distance between the target and the radar;

[0067] In other words, the distance between the moving object and the radar transmission center can be determined by the phase and timing of the echo signal.

[0068] (f) Radar Angles: These are the angles between the radar and the target, including azimuth and elevation angles, which are very important in target localization and tracking; in this embodiment, the radar sensor is set to measure angles of 45°.

[0069] S102: Based on the same radar data corresponding to different detection frequencies, the same speed change state, and the same distance for each type of moving object, plot curves of the same radar data for automobiles and other types of moving objects in a preset coordinate system. The horizontal axis of the coordinate system is the radar detection frequency, and the vertical axis is the same radar data.

[0070] That is, by using the controlled variable method, we plotted the Doppler frequency shift, polarization state, and echo signal amplitude variation curves of automobiles and other types of moving objects at different radar detection frequencies.

[0071] For example, such as Figure 3 As shown, Figure 3 The curves show the variation of echo signal amplitudes for vehicle-type targets and non-vehicle-type targets at different radar detection frequencies. The curve with a larger peak value represents the signal spectrum curve when the moving object in the detection area is a vehicle, while the curve with a smaller peak value represents the signal spectrum curve when the moving object in the detection area is not a vehicle (including cases where the moving object is another type of object or no moving object is detected in the detection area).

[0072] S103: Based on the curves of the same radar data corresponding to the car and other types of moving objects, determine whether the same radar data is valid, that is, determine whether there is a significant difference between the two curves. If so, it means that the radar data can distinguish the car and other types of moving objects to a certain extent, and proceed to step S104.

[0073] This step focuses on comparing and analyzing the differences in image features between the curves corresponding to small and medium-sized cars and the curves corresponding to other types of moving objects, so as to realize the potential recognition feature differences between small and medium-sized cars and other types of moving objects based on a single-factor perspective.

[0074] Specifically, if a discrimination threshold curve can be determined based on the curves of the same radar data corresponding to the vehicle and other types of moving objects, then the same radar data is considered valid. In this embodiment, the discrimination threshold curve needs to meet the following conditions:

[0075] 1) The discrimination threshold curve lies between the curves corresponding to the same radar data for the vehicle and other types of moving objects, that is, between the vehicle type curve and the non-vehicle type curve, which can effectively distinguish vehicle type targets from non-vehicle type targets, such as... Figure 3 As shown. Furthermore, in some cases, since the two curves may intersect, it is sufficient that a certain portion (a preset first threshold, such as 70%) of the discrimination threshold curve lies between the two curves;

[0076] 2) The discrimination threshold curve has an effective detection range. The accuracy of identifying the type of moving object based on the threshold point corresponding to the effective detection range on the discrimination threshold curve reaches 60%. That is, for any detection frequency within the effective detection range, if the same radar data of the moving object is collected multiple times at that detection frequency, the probability of correctly identifying the type of moving object based on the collected radar data and the threshold corresponding to the detection frequency on the discrimination threshold curve reaches 60%.

[0077] Furthermore, the morphological characteristics and / or extreme value characteristics of the curves corresponding to the same radar data for the vehicle and other types of moving objects can also be used to determine whether there is a difference between the two curves. For example, in an optional embodiment, the radar data difference for the vehicle and other types of moving objects at the same detection frequency is calculated separately to obtain the radar data difference corresponding to each detection frequency. If the radar data difference corresponding to more than a certain proportion of detection frequencies is greater than a preset difference threshold, then the two curves are considered to have a difference. Figure 3 Taking the echo signal amplitude curve as an example, within the detection frequency range of 0-50, if the signal spectrum difference corresponding to more than 20% of the detection frequency is greater than 30, it is considered that there is a difference in the change curve of the echo signal amplitude between the car type target and the non-car type target, that is, the echo signal amplitude is considered to be valid.

[0078] S104: Use the same radar data as sample radar data. That is, use all valid radar data as sample data for subsequent training of the support vector machine model to achieve multi-factor joint analysis.

[0079] Furthermore, in another optional embodiment, when determining the validity of the same radar data based on the curves corresponding to the vehicle and other types of moving objects, the optimal detection frequency corresponding to the same radar data is also determined based on the morphological and extreme value characteristics of these two curves. The optimal detection frequency can clearly distinguish the detection frequency of the signal spectrum of the same radar data for vehicles and other types of moving objects. For example, as... Figure 3 As shown, when the detection frequency is 15, the signal spectrum of the car differs significantly from the signal spectra of other types of moving objects. There are various ways to determine the optimal detection frequency. In this embodiment, the signal spectrum difference between the car and other types of moving objects at the same detection frequency is calculated, and the detection frequency corresponding to the maximum value of the signal spectrum difference is selected as the optimal detection frequency.

[0080] Then, based on the optimal detection frequency corresponding to the same radar data, the same radar data corresponding to different speed changes and different distances for each type of moving object is collected as sample radar data.

[0081] By determining the optimal detection frequency for effective radar data and then re-collecting radar data at that optimal frequency as sample data for subsequent model training, the accuracy of model recognition can be improved.

[0082] S105: Generate a standard training sample set based on the attribute information of each type of moving object and the corresponding sample radar data of different speed changes and different distances.

[0083] Single-factor analysis can only initially reveal potential local identification features, without comprehensively considering the impact of the interaction of various factors on the overall effective identification. To further conduct multi-factor linkage identification analysis, this embodiment normalizes the effective radar data to generate a standard training sample set. The normalized sample data format is shown in Table 1.

[0084] Table 1: Standardized Sample Data Format

[0085]

[0086] S106: Based on the standard training sample set, train the preset machine learning model to obtain a trained machine learning model, wherein the machine learning model is a support vector machine model.

[0087] Due to the relatively small sample size, this embodiment employs a Support Vector Machine (SVM) model. The model is trained using a standard training sample set, and a ten-fold cross-validation method is used to enhance its generalization ability. Based on classification accuracy, the correlation of different feature information and the weight ratio of each factor in target recognition are statistically analyzed to achieve preliminary multi-factor joint feature recognition analysis of moving object categories. During this feature analysis process, relevant characteristics are combined to form a training sample set, and experimental tests are conducted based on the SVM algorithm. The prediction model accuracy of different feature combinations is ranked from highest to lowest, analyzing the impact of different features on the overall recognition accuracy of the model. This process eliminates invalid recognition features, simplifies the parameters of effective recognition features, and obtains the optimal combination recognition characteristics. This achieves preliminary category recognition based on microwave radar technology, reducing the probability of invalid vehicle recognition by the camera in abnormal environments.

[0088] S107: The radar data of the target to be tested is acquired by the microwave radar sensor, and the target to be tested is initially classified by the trained machine learning model based on the radar data of the target to be tested, so as to obtain the preliminary identification result.

[0089] In this step, the collected radar data are those deemed valid in the previous steps. In this embodiment, Doppler frequency shift, polarization state, and echo signal amplitude are all deemed valid; therefore, in practical applications, these three types of radar data are collected for the target. Further, in another optional embodiment, each radar data point can be collected separately according to the previously determined optimal detection frequency. Then, a preliminary identification result is obtained using a trained support vector machine model.

[0090] For the second part, in this embodiment, the deep learning model adopts the YOLOv5 deep learning algorithm framework, such as... Figure 4As shown, its network structure includes an input, a backbone, a neck, and an output. The backbone extracts image features, the neck fuses features, and the output head predicts the target's location and category. Yolov5 employs an anchor-based object detection method, using anchor boxes to generate prior boxes to predict the target's location and size, while using algorithms such as DIOU_Loss and DIOU_nms to handle overlapping, repetitive, and uncertain cases.

[0091] like Figure 5 As shown, the second part specifically includes the following steps:

[0092] S201: Acquire captured images using a camera device, and perform data cleaning on the captured images to obtain a sample dataset.

[0093] Specifically, the high-definition camera configured on the curb machine captures images of actual roadside scenes and performs data cleaning.

[0094] S202: Based on the preset training image size, perform data augmentation and adaptive anchor box calculation on the sample dataset, and scale, fill, annotate and convert the images in the sample dataset to obtain a standard image sample set.

[0095] The image size captured by the high-definition camera is 1920×1080 pixels, while the training image size of the YOLOv5 model is generally 640×640 pixels. Therefore, it is necessary to perform Mosaic data augmentation and adaptive anchor box calculation on the sample dataset, and perform operations such as scaling, padding, annotation and format conversion on the images to obtain a standard image sample set. Among them, the format conversion is to convert the image format captured by the high-definition camera to the image input format required by the YOLOv5 model.

[0096] S203: Based on the standard image sample set and preset training hyperparameters, train the preset deep learning model to obtain the trained deep learning model.

[0097] Specifically, the pre-trained weight file of the YOLOv5 model was obtained as the initial network parameters, and the main initial hyperparameters were set as follows: batch size = 32, image input size = 640, training batches = 100, initial learning rate lr0 = 0.01, cosine annealing hyperparameter (learning rate descent parameter) lrf = 0.01, learning rate momentum = 0.937, weight decay coefficient = 0.0005, and classification loss coefficient cls = 0.5. The CIOU_Loss function was used as the loss function.

[0098] During training, a fitting learning rate is used for training. Training is stopped once a certain validation accuracy (e.g., 98%) is achieved. By monitoring changes in model accuracy during training, the model and parameters are adjusted based on the results to balance the issues of model overfitting and model training accuracy.

[0099] The model is optimized by repeatedly training it, adjusting hyperparameters and network structure, and finally saving the trained YOLOv5 model.

[0100] Furthermore, after model training is complete, the trained model can be evaluated using recall, precision, and F1 score, calculated as follows:

[0101] Recall = TP / (TP+FN);

[0102] Precision = TP / (TP+FP);

[0103] F1=(2×Precision×Recall) / (Precision+Recall);

[0104] In this context, TP (True Positive) represents the number of positive samples identified as positive by the model; FP (False Positive) represents the number of positive samples identified as negative by the model; and FN (False Negative) represents the number of negative samples identified as negative by the model.

[0105] S204: When the initial identification result is a car, the image of the target to be tested is captured by a camera device, and the image of the target to be tested is identified by a pre-trained deep learning model to obtain the category of the target to be tested.

[0106] That is, if the preliminary identification result obtained in step S107 is a car, the image of the target to be tested is captured by the high-definition camera of the curb machine, and then the category of the target to be tested is identified by the trained YOLOv5 model.

[0107] This embodiment fully utilizes the inherent information acquired by the microwave radar equipment, effectively improving the utilization of data information and expanding the potential of microwave radar in the field of object recognition to a certain extent. Furthermore, it effectively reduces the probability of invalid captures by the curb-mounted camera in real-world scenarios, enhances its resistance to interference from multiple environmental variables, and effectively solves problems such as capturing empty cars in parking spaces, capturing passing vehicles, and mistakenly capturing non-small and medium-sized cars due to interference from equipment and environmental factors in real-world scenarios.

[0108] Example 2

[0109] This embodiment is a computer-readable storage medium corresponding to the above embodiments, on which a computer program is stored. When the program is executed by a processor, it implements the various steps of the parking vehicle identification method based on microwave radar assisted in the above embodiments and can achieve the same technical effect, which will not be repeated here.

[0110] In summary, this invention provides a microwave radar-assisted parking vehicle identification method and storage medium. By collecting radar data of the target, preliminary category identification is performed based on the radar data. Capture and identification are only triggered when the preliminary identification result is a car. This effectively avoids problems such as capturing empty vehicles, capturing passing vehicles, and mistakenly capturing non-small and medium-sized vehicles due to equipment and environmental interference, reducing the probability of invalid captures, enhancing the anti-interference capability against multiple environmental variables, and achieving intelligent identification of small and medium-sized vehicles with parking intentions. Simultaneously, it fully utilizes the inherent information acquired by the microwave radar equipment, effectively improving the utilization of data information and expanding the potential possibilities of microwave radar in the field of object recognition to a certain extent. This invention can effectively reduce the probability of invalid captures by cameras in parking spaces and improve the accuracy of parking vehicle identification.

[0111] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for identifying parked vehicles based on microwave radar assistance, characterized in that, include: Radar data of the target under test is acquired by a microwave radar sensor, and the target under test is initially classified based on the radar data to obtain a preliminary identification result. If the preliminary identification result is a car, then an image of the target to be tested is captured by a camera device, and the image of the target to be tested is identified by a pre-trained deep learning model to obtain the category of the target to be tested; Before acquiring radar data of the target under test through a microwave radar sensor and performing preliminary category identification of the target under test based on the radar data to obtain a preliminary identification result, the process further includes: The system acquires attribute information for various types of moving objects and, based on the same test environment and movement path, collects radar data for each type of moving object at different detection frequencies, speed change states, and distances using a microwave radar sensor. The types of moving objects include cars, electric vehicles, and people. The attribute information includes length, width, and height. The speed change states include uniform movement and deceleration to a stop. The distance is the distance between the moving object and the transmission center of the microwave radar sensor. The radar data includes Doppler frequency shift, polarization state, and echo signal amplitude. Based on the same radar data corresponding to different detection frequencies, the same speed change state, and the same distance for each type of moving object, curves of automobiles and other types of moving objects corresponding to the same radar data are plotted in a preset coordinate system. The horizontal axis of the coordinate system is the detection frequency, and the vertical axis is the same radar data. Based on the curves of the same radar data corresponding to the car and other types of moving objects, determine whether the same radar data is valid; If valid, the same radar data will be used as sample radar data. A standard training sample set is generated based on the attribute information of each type of moving object and the corresponding sample radar data of different speed changes and different distances. The preset machine learning model is trained based on the standard training sample set to obtain a trained machine learning model, wherein the machine learning model is a support vector machine model. The deep learning model is a YOLOv5 model, which includes an input, a backbone network, a neck network, and an output. The input terminal is used to input images; The backbone network is used to extract image features; The neck network is used to fuse image features; The output is used to predict the location and category of the target.

2. The parking vehicle identification method based on microwave radar assistance according to claim 1, characterized in that, The step of determining whether the same radar data is valid based on the curves corresponding to the same radar data for the vehicle and other types of moving objects is as follows: If a discrimination threshold curve exists corresponding to the same radar data, then the same radar data is determined to be valid. The portion of the discrimination threshold curve above the preset first threshold is located between the curves corresponding to the same radar data for the vehicle and other types of moving objects, and the discrimination threshold curve has an effective detection range. The recognition accuracy of moving object type identification based on the threshold point corresponding to the effective detection range on the discrimination threshold curve reaches the preset second threshold.

3. The parking vehicle identification method based on microwave radar assistance according to claim 1, characterized in that, The specific steps for using the same radar data as sample radar data are as follows: Based on the morphological and extreme value characteristics of the curves corresponding to the same radar data for the vehicle and other types of moving objects, the optimal detection frequency corresponding to the same radar data is determined. Based on the optimal detection frequency corresponding to the same radar data, the same radar data corresponding to different speed changes and different distances for each type of moving object is collected as sample radar data.

4. The parking vehicle identification method based on microwave radar assistance according to claim 1, characterized in that, The preliminary category identification of the target under test based on the radar data, to obtain a preliminary identification result, is specifically as follows: Based on the radar data of the target to be tested, the trained machine learning model is used to perform preliminary category identification of the target to be tested, and a preliminary identification result is obtained.

5. The parking vehicle identification method based on microwave radar assistance according to claim 1, characterized in that, Before identifying the image of the target object using a pre-trained deep learning model to obtain the category of the target object, the method further includes: The captured images are obtained by using a camera device, and the captured images are cleaned to obtain a sample dataset; Based on the preset training image size, data augmentation and adaptive anchor box calculation are performed on the sample dataset, and the images in the sample dataset are scaled, filled, labeled and format converted to obtain a standard image sample set. Based on the standard image sample set and preset training hyperparameters, a preset deep learning model is trained to obtain a trained deep learning model.

6. The parking vehicle identification method based on microwave radar assistance according to claim 1, characterized in that, The loss function of the YOLOv5 model is the CIOU_Loss function.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Images