Vehicle classification method and system
By matching and clustering point cloud data using extended Kalman filtering and the DBSCAN algorithm, the problem of millimeter-wave radar's difficulty in distinguishing between large and small vehicles in vehicle target tracking and classification is solved, achieving a simple and computationally inexpensive vehicle size classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUZHOU RUIXIN GUANYUAN TERAHERTZ TECH CO LTD
- Filing Date
- 2022-09-23
- Publication Date
- 2026-06-26
AI Technical Summary
Millimeter-wave radar has difficulty distinguishing between large and small vehicles in vehicle target tracking and classification.
The extended Kalman filter algorithm and the DBSCAN algorithm are used to match and cluster point cloud data, and the vehicle size is classified by the maximum distance between the point cloud data of the target vehicle.
It achieves a simple and computationally inexpensive vehicle size classification system, and has good application prospects.
Smart Images
Figure CN115439648B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle inspection technology, specifically to a vehicle classification method and system. Background Technology
[0002] The increasingly complex road traffic environment is driving my country's development towards intelligent transportation, in which millimeter-wave radar is playing an increasingly important role. In the current transportation field, millimeter-wave radar has significant advantages in measuring vehicle distance, speed, and angle. Frequency-modulated continuous wave (FMCW) millimeter-wave radar technology measures parameters such as distance, speed, and angle of surrounding objects by transmitting a series of continuously frequency-modulated millimeter waves through an antenna and receiving the reflected signals from targets.
[0003] Millimeter-wave radar has advantages such as being able to operate in any lighting and weather conditions, quickly acquiring the speed of the vehicle being measured, being small and lightweight, and being less expensive than lidar. However, it also has some drawbacks, such as the inability to distinguish between large and small vehicles in vehicle target tracking and classification.
[0004] Therefore, in order to address the aforementioned technical problems, it is necessary to provide a vehicle classification method and system. Summary of the Invention
[0005] In view of this, the purpose of the present invention is to provide a vehicle classification method and system to achieve the classification of target vehicles by size.
[0006] To achieve the above objectives, an embodiment of the present invention provides the following technical solution:
[0007] A vehicle classification method, the method comprising:
[0008] S1. Acquire multiple frames of real-time data of the observation area, and obtain the corresponding point cloud dataset D based on the real-time data, where D = {d1, d2, ..., d...} m};
[0009] S2. Match the point cloud data in the point cloud dataset D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to the multiple successfully matched point cloud data to the target vehicle set O.
[0010] S3. Perform cluster analysis on the unmatched point cloud data in step S2 using the DBSCAN algorithm, and add the target vehicles corresponding to multiple point cloud data that exceed the matching threshold to the target vehicle set O.
[0011] S4. Based on the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud dataset D, obtain the distance between any two point cloud data and take the maximum distance as the vehicle length.
[0012] S5. Based on the vehicle length and the threshold value, assign large and small vehicle tags to the vehicles. Assign large vehicle tags to target vehicles whose length is greater than or equal to the threshold value, and assign small vehicle tags to target vehicles whose length is less than the threshold value.
[0013] In one embodiment, step S1 specifically includes:
[0014] Multiple frames of real-time data of the observation area were acquired using a millimeter-wave radar based on frequency-modulated continuous wave.
[0015] Filter and denoise real-time data;
[0016] For the filtered and denoised real-time data, the corresponding point cloud dataset D is obtained according to the principle of frequency-modulated continuous wave.
[0017] In one embodiment, each point cloud data in the point cloud dataset D includes velocity information v, distance information r, and angle information a.
[0018] In one embodiment, the extended Kalman filter algorithm prediction in step S2 includes:
[0019] Linearize the state transition equation and the observation equation;
[0020] Prediction is performed using the Kalman filter algorithm based on the linearized state transition equation and observation equation.
[0021] In one embodiment, step S3 is followed by:
[0022] The point cloud data corresponding to all target vehicles is converted into two dimensions to obtain the centroid position (x, y) and velocity (v) of each target vehicle. x v y ).
[0023] In one embodiment, step S3 is followed by:
[0024] Based on the constant acceleration model and the centroid position and velocity of the target vehicle, predict the centroid position and velocity of the target vehicle in the subsequent frame point cloud dataset D;
[0025] Compare whether a new target vehicle that did not appear in the previous frame's point cloud dataset D is matched in a certain frame's point cloud dataset D;
[0026] If a new target vehicle is matched, the centroid position and velocity of the new target vehicle are compared with those of one of the predicted target vehicles within the preset distance threshold and preset velocity threshold. If they are, the new target vehicle is deleted and the point cloud data corresponding to the new target vehicle is mapped to the matched target vehicle. If not, the new target vehicle is added to the target vehicle set 0.
[0027] If no new target vehicle is matched, the target vehicle set O remains unchanged.
[0028] In one embodiment, step S3 is followed by:
[0029] A counting label is configured for each target vehicle. The initial value of the counting label is 0. When the target vehicle disappears in a certain frame of point cloud dataset D, the counting label is incremented by 1.
[0030] When the number of count tags exceeds the preset tag threshold, the target vehicle is removed from the target vehicle set O.
[0031] Once the target vehicle is re-matched in a point cloud dataset D of a certain frame, the count label is set to 0.
[0032] In one embodiment, step S5 further includes:
[0033] Traverse all frame point cloud datasets D and assign a car label to the first matched target vehicle.
[0034] For target vehicles whose vehicle length is greater than or equal to a threshold value in x consecutive frames, they are locked as large vehicle labels, where x≥2;
[0035] For target vehicles whose vehicle length is less than a threshold value in consecutive y frames, they are locked as car labels, where y≥2.
[0036] In one embodiment, step S3 is followed by:
[0037] Point cloud data exceeding the matching threshold are identified as noise and discarded.
[0038] Another embodiment of the present invention provides the following technical solution:
[0039] A vehicle classification system, the system comprising:
[0040] The data acquisition module is used to acquire multiple frames of real-time data from the observation area and obtain the corresponding point cloud dataset D based on the real-time data, where D = {d1, d2, ..., d...} m};
[0041] The data matching unit is used to match the point cloud data in the point cloud dataset D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to multiple successfully matched point cloud data to the target vehicle set O; and to perform cluster analysis on the unmatched point cloud data using the DBSCAN algorithm, and add the target vehicles corresponding to multiple point cloud data exceeding the matching threshold to the target vehicle set O.
[0042] The length acquisition unit is used to obtain the distance between any two point cloud data based on the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud dataset D, and take the maximum distance as the vehicle length.
[0043] The tag configuration unit is used to configure large and small vehicle tags for vehicles based on the vehicle length and threshold value. For target vehicles whose length is greater than or equal to the threshold value, a large vehicle tag is configured, and for target vehicles whose length is less than the threshold value, a small vehicle tag is configured.
[0044] The present invention has the following beneficial effects:
[0045] This invention uses the extended Kalman filter algorithm and the DBSCAN algorithm to match point cloud data to obtain target vehicles, and classifies vehicle size by the maximum distance between the point cloud data corresponding to each target vehicle. It has the characteristics of low computational load and relatively simple classification, and has good application prospects. Attached Figure Description
[0046] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0047] Figure 1 This is a flowchart illustrating the vehicle classification method in this invention;
[0048] Figure 2 This is a schematic diagram of the vehicle classification system in this invention;
[0049] Figure 3 This is a schematic diagram illustrating the calculation principle of vehicle length in Embodiment 1 of the present invention;
[0050] Figure 4 This is a schematic diagram illustrating the principle of point cloud data 2D conversion in Embodiment 2 of the present invention;
[0051] Figure 5 This is a schematic diagram showing the position and velocity of the center of mass in Embodiment 2 of the present invention;
[0052] Figure 6 This is a schematic diagram of frame skipping for the target vehicle in Embodiment 2 of the present invention. Detailed Implementation
[0053] To enable those skilled in the art to better understand the technical solutions in this invention, the following will clearly and completely describe the technical solutions in the embodiments of this invention in conjunction with the accompanying drawings in the embodiments of this invention. Apparently, the described embodiments are only a part of the embodiments of this invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this invention without creative efforts shall fall within the protection scope of this invention.
[0054] As shown Figure 1 in the figure, this invention discloses a vehicle classification method, including:
[0055] S1. Obtain multiple frames of real-time data in the observation area, and obtain the corresponding point cloud data set D according to the real-time data, where D = {d1, d2,..., d m};
[0056] S2. Match the point cloud data in the point cloud data set D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to the successfully matched multiple point cloud data to the target vehicle set O;
[0057] S3. Perform clustering analysis on the point cloud data that is not successfully matched in step S2 using the DBSCAN algorithm, and add the target vehicles corresponding to the multiple point cloud data exceeding the matching threshold to the target vehicle set O;
[0058] S4. According to the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud data set D, obtain the distance between any two point cloud data, and take the maximum distance as the vehicle length;
[0059] S5. Configure large and small vehicle labels for the vehicle according to the size of the vehicle length and the threshold value. Configure a large vehicle label for the target vehicle whose vehicle length is greater than or equal to the threshold value, and configure a small vehicle label for the target vehicle whose vehicle length is less than the threshold value.
[0060] As shown Figure 2 in the figure, this invention also discloses a vehicle classification system, including:
[0061] A data acquisition module 10, configured to obtain multiple frames of real-time data in the observation area, and obtain the corresponding point cloud data set D according to the real-time data, where D = {d1, d2,..., d m};
[0062] The data matching unit 20 is used to match the point cloud data in the point cloud dataset D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to multiple successfully matched point cloud data to the target vehicle set O; and to perform cluster analysis on the unmatched point cloud data using the DBSCAN algorithm, and add the target vehicles corresponding to multiple point cloud data exceeding the matching threshold to the target vehicle set O.
[0063] The length acquisition unit 30 is used to obtain the distance between any two point cloud data based on the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud dataset D, and take the maximum distance as the vehicle length.
[0064] The tag configuration unit 40 is used to configure large and small vehicle tags for vehicles based on the vehicle length and threshold value. For target vehicles whose length is greater than or equal to the threshold value, a large vehicle tag is configured, and for target vehicles whose length is less than the threshold value, a small vehicle tag is configured.
[0065] The present invention will be further described below with reference to specific embodiments.
[0066] Example 1:
[0067] In this embodiment, the obtained point cloud data is classified by a tracking algorithm, the maximum spatial length of a single target vehicle is calculated, and the result is compared with a set threshold value to classify the target vehicle.
[0068] Specifically, the vehicle classification method in this embodiment includes the following steps:
[0069] S1. Acquire multiple frames of real-time data of the observation area, and obtain the corresponding point cloud dataset D based on the real-time data, where D = {d1, d2, ..., d...} m}
[0070] Specifically, in this step, a series of continuously frequency-modulated millimeter waves is transmitted outward by a millimeter-wave radar based on frequency-modulated continuous waves, and the reflected signal from the target is received, thereby obtaining multiple frames of real-time data of the observation area (i.e., real-time data acquired by the millimeter-wave radar sensor), and the real-time data is filtered and denoised.
[0071] Then, based on the frequency modulated continuous wave (FMCW) principle, the corresponding point cloud dataset D, D={d1, d2, ..., d...}, is obtained from the filtered and denoised real-time data. m}, d1~d m The data consists of point cloud data, each point cloud data includes velocity information v, distance information r, and angle information a.
[0072] It should be understood that the frequency modulated continuous wave (FMCW) principle mentioned in this embodiment is a conventional principle in the prior art, and will not be elaborated here.
[0073] S2. Match the point cloud data in the point cloud dataset D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to the multiple successfully matched point cloud data to the target vehicle set O.
[0074] In this step, the point cloud dataset D is matched using the Extended Kalman Filter (EKF) algorithm. The Kalman Filter algorithm is a well-established technique and comprises five formulas: the formula for calculating the prior state estimate, the formula for calculating the prior error covariance, the formula for calculating the correction matrix, the formula for updating the observations, and the formula for updating the error covariance. Existing Kalman Filter algorithms predict based on linear state transition equations and observation equations. However, the state transition equations and observation equations in this step are nonlinear. Therefore, this step uses the EKF algorithm to predict the target vehicle. First, the state transition equations and observation equations are linearized, and then the Kalman Filter algorithm is used for prediction based on the linearized state transition equations and observation equations.
[0075] Furthermore, during the prediction process using the Extended Kalman Filter (EKF) algorithm, the parameters in the EKF algorithm are updated, including the Kalman gain, state matrix, and covariance matrix. This is a standard technique in the Kalman Filter algorithm and will not be elaborated upon here.
[0076] When the point cloud data in the point cloud dataset D is successfully matched with the target vehicle data predicted by the extended Kalman filter algorithm, the target vehicles corresponding to the multiple successfully matched point cloud data are added to the target vehicle set O, finally obtaining the target vehicle set O = {o1, o2, ..., o...} n1 That is, based on the extended Kalman filter algorithm, n1 target vehicles are matched.
[0077] S3. Perform cluster analysis on the unmatched point cloud data in step S2 using the DBSCAN algorithm, and add the target vehicles corresponding to multiple point cloud data that exceed the matching threshold to the target vehicle set O.
[0078] In this step, the DBSCAN algorithm is used to re-match the point cloud data that failed to match in step S2 using the extended Kalman filter algorithm. The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is a density-based clustering algorithm. DBSCAN defines a cluster as the largest set of density-connected points, which can divide regions with sufficiently high density into clusters and can discover clusters of arbitrary shapes in noisy spatial databases.
[0079] When the point cloud data in the point cloud data set D exceeds the matching threshold obtained by the DBSCAN algorithm, multiple point cloud data that exceed the matching threshold are matched into a new target vehicle, and the target vehicle is added to the target vehicle set O. Finally, the target vehicle set O = {o1, o2, …, o n1 ,o n1+1 ,o n1+2 ,…,o n1+n2}, that is, n2 target vehicles are matched based on the DBSCAN algorithm.
[0080] Preferably, after the point cloud data matching in the above two steps, the point cloud data matched by both the extended Kalman filter algorithm and the DBSCAN algorithm are determined as noise points and discarded.
[0081] S4. According to the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud data set D, obtain the distance between any two point cloud data, and take the maximum distance as the vehicle length.
[0082] In each frame of the point cloud data set D in this embodiment, each target vehicle corresponds to at least two point cloud data. According to the distance information r and angle information a corresponding to the two point cloud data, the distance between the two point cloud data is calculated using the cosine theorem.
[0083] As shown in the reference Figure 3 , the angle information corresponding to the two point cloud data is a1 and a2 respectively, and the corresponding distance information is r1 and r2 respectively. Then, the distance CarLength between the two point cloud data is:
[0084] CarLength =
[0085] Calculate the distance between any two point cloud data corresponding to each target vehicle through the above formula, and the maximum distance CarLength max is the vehicle length.
[0086] S5. According to the size relationship between the vehicle length and the threshold value, configure large and small vehicle tags for the vehicle. Configure a large vehicle tag (BigCarTag) for the target vehicle whose vehicle length is greater than or equal to the threshold value, and configure a small vehicle tag (SmallCarTag) for the target vehicle whose vehicle length is less than the threshold value.
[0087] Furthermore, in view of the problem that the vehicle length will repeatedly affect the judgment of the large and small vehicle tags of the target vehicle, a certain threshold can also be set for tag locking. Specifically:
[0088] Traverse all frames of the point cloud data set D, and configure a small vehicle tag (SmallCarTag) for each target vehicle, that is, the initial tag of the target vehicle is the small vehicle tag (SmallCarTag);
[0089] For target vehicles with a vehicle length greater than or equal to the threshold value in consecutive x frames, lock them as the big vehicle label (BigCarTag), where x ≥ 2;
[0090] For target vehicles with a vehicle length less than the threshold value in consecutive y frames, lock them as the small vehicle label (SmallCarTag), where y ≥ 2.
[0091] In this embodiment, when x = 3 and y = 5, after determining a target vehicle as a big vehicle continuously for 3 times, lock the label of this target vehicle as the big vehicle label (BigCarTag), and after determining a target vehicle as a small vehicle continuously for 5 times, lock the label of this target vehicle as the small vehicle label (SmallCarTag). In subsequent large and small vehicle classifications, maintain the locked label unchanged, and no longer perform large and small vehicle classifications on the target vehicle with the locked label.
[0092] Embodiment 2:
[0093] The vehicle classification method in this embodiment is similar to that in Embodiment 1. The difference is that the target vehicle may have a frame skipping situation where it suddenly disappears. For the frame skipping situation, first, two-dimensionalize the point cloud data corresponding to all target vehicles to obtain the centroid position (x, y) and speed (v x , v y ) of each target vehicle.
[0094] Refer Figure 4 The schematic diagram of the two-dimensionalization of the point cloud data is shown. The height of the millimeter-wave radar is h, the distance in the point cloud data is r, the angle is a, and the distance after the two-dimensionalization of the point cloud data is r', satisfying r' = r * sina.
[0095] Refer Figure 5 The schematic diagram of the centroid position and speed is shown. Taking four point cloud data corresponding to a target vehicle as an example, two-dimensionalize the 4 point cloud data, and the centroid position (x, y) and speed (v x , v y ) of this target vehicle can be obtained according to the two-dimensionalized point cloud data.
[0096] For the frame skipping situation of the target vehicle, the processing method is as follows:
[0097] According to the constant acceleration model and the centroid position and speed of the target vehicle, predict the centroid position and speed corresponding to the target vehicle in the subsequent frame point cloud dataset D;
[0098] Compare whether a new target vehicle that did not appear in the previous frame point cloud dataset D is matched in a certain frame point cloud dataset D;
[0099] If a new target vehicle is matched, compare whether the centroid position and speed of the new target vehicle are within the preset distance threshold and preset speed threshold of the centroid position and speed of one of the predicted target vehicles. If so, delete the new target vehicle and correspond the point cloud data corresponding to the new target vehicle to the matched target vehicle. If not, add the new target vehicle to the target vehicle set 0;
[0100] If no new target vehicle is matched, keep the target vehicle set O unchanged.
[0101] Refer Figure 6 As shown, taking three frames of data as an example for illustration, through the scheme in Embodiment 1, target vehicles o1, o2, and o3 can be matched in the first frame of data, only target vehicles o1 and o2 can be matched in the second frame of data, and in addition to target vehicles o1 and o2, a new target vehicle o4 outside o1 and o2 is matched in the third frame of data through the DBSCAN algorithm. After comparing the centroid position and speed, it is found that the centroid position and speed of the new target vehicle o4 are within the preset distance threshold and preset speed threshold of the centroid position and speed of the target vehicle o3 in the first frame of data, indicating that in this case, the target vehicle o3 has a frame skipping situation (i.e., skipping from the first frame to the third frame and not appearing in the second frame). The new target vehicle o4 in the third frame is the target vehicle o3 in the first frame. At this time, delete the new target vehicle o4 and correspond the point cloud data corresponding to the new target vehicle o4 to the target vehicle o3.
[0102] At this time, the target vehicle set O in the first frame of data = {o1, o2, o3}. Assume that the target vehicle is deleted from the target vehicle set O when it disappears for 3 frames. The target vehicle set O in the second frame of data = {o1, o2, o3}. In the third frame of data, the target vehicle o3 still does not disappear for 3 frames or more and will not be deleted from the target vehicle set. Moreover, the new target vehicle o4 is the target vehicle o3 that disappeared in the second frame, so the new target vehicle o4 will not be added to the target vehicle set. The target vehicle set O in the third frame = {o1, o2, o3}.
[0103] In another case, if the centroid position and speed of the new target vehicle o4 are not within the preset distance threshold and preset speed threshold of the centroid position and speed of the target vehicle o3 in the first frame of data, it means that the new target vehicle o4 is the target vehicle that first appears in the third frame of data and is different from the previous target vehicles o1, o2, and o3. At this time, add the new target vehicle o4 to the target vehicle set O.
[0104] At this point, the target vehicle set O = {o1, o2, o3} in the first frame of data is assumed to be deleted from the target vehicle set O after the target vehicle disappears for 3 frames. The target vehicle set O = {o1, o2, o3} in the second frame of data is also assumed to be {o1, o2, o3}. In the third frame of data, the target vehicle o3 has not disappeared for 3 frames or more and will not be deleted from the target vehicle set. The target vehicle set O = {o1, o2, o3, o4} in the third frame is also assumed to be {o1, o2, o3, o4}.
[0105] In another scenario, if the newly generated point cloud data cannot be matched with a new target vehicle using the DBSCAN algorithm, the newly generated point cloud data is considered to be noisy data, and the noisy data can be deleted.
[0106] Example 3:
[0107] The vehicle classification method in this embodiment is similar to that in embodiment 1. The difference is that the target vehicle may leave the observation area. If the target vehicle leaves the observation area in a certain frame, the target vehicle cannot be matched in that frame.
[0108] The specific handling method for situations where the target vehicle leaves the observation area is as follows:
[0109] For each target vehicle, a counter label OutCount is configured. The initial value of the counter label OutCount is 0. When the target vehicle disappears in a certain frame of point cloud dataset D, the counter label OutCount is incremented by 1.
[0110] When the count tag OutCount is greater than the preset tag threshold, the target vehicle is removed from the target vehicle set O;
[0111] Once the target vehicle is re-matched in a point cloud dataset D of a certain frame, the count label OutCount is set to 0.
[0112] Specifically, if a target vehicle leaves the observation area for the first time in a frame, the target vehicle's count label OutCount is updated to 1. If the target vehicle returns to the observation area in the next frame, the count label OutCount is set to 0 and the count is restarted. If the target vehicle is still outside the observation area in the next frame, the target vehicle's count label OutCount is updated to 2, and so on.
[0113] In this embodiment, the preset tag threshold is set to 3. When the detected count tag OutCount is greater than 3, the target vehicle is removed from the target vehicle set O.
[0114] As can be seen from the above technical solutions, the present invention has the following advantages:
[0115] This invention uses the extended Kalman filter algorithm and the DBSCAN algorithm to match point cloud data to obtain target vehicles, and classifies vehicle size by the maximum distance between the point cloud data corresponding to each target vehicle. It has the characteristics of low computational load and relatively simple classification, and has good application prospects.
[0116] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0117] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A vehicle classification method, characterized in that, The method includes: S1. Acquire multiple frames of real-time data of the observation area, and obtain the corresponding point cloud dataset D based on the real-time data, where D = {d1, d2, ..., d...} m }, d1~d m The data consists of point cloud data, each point cloud data includes velocity information v, distance information r, and angle information a; S2. Match the point cloud data in the point cloud dataset D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to the multiple successfully matched point cloud data to the target vehicle set O. S3. Perform cluster analysis on the unmatched point cloud data in step S2 using the DBSCAN algorithm, and add the target vehicles corresponding to multiple point cloud data that exceed the matching threshold to the target vehicle set O. S4. Based on the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud dataset D, obtain the distance between any two point cloud data and take the maximum distance as the vehicle length. S5. Based on the vehicle length and the threshold value, assign large and small vehicle tags to the vehicles. Assign a large vehicle tag to target vehicles whose length is greater than or equal to the threshold value, and assign a small vehicle tag to target vehicles whose length is less than the threshold value. The step S3 is followed by: The point cloud data corresponding to all target vehicles is converted into two dimensions to obtain the centroid position (x, y) and velocity (v) of each target vehicle. x v y ); Based on the constant acceleration model and the centroid position and velocity of the target vehicle, predict the centroid position and velocity of the target vehicle in the subsequent frame point cloud dataset D; Compare whether a new target vehicle that did not appear in the previous frame's point cloud dataset D is matched in a certain frame's point cloud dataset D; If a new target vehicle is matched, the centroid position and velocity of the new target vehicle are compared with those of one of the predicted target vehicles. If they are within the range of a preset distance threshold and a preset velocity threshold, the new target vehicle is deleted and the point cloud data corresponding to the new target vehicle is mapped to the matched target vehicle. If not, the new target vehicle is added to the target vehicle set O. If no new target vehicle is matched, the target vehicle set O remains unchanged.
2. The vehicle classification method according to claim 1, characterized in that, Step S1 specifically involves: Multiple frames of real-time data of the observation area were acquired using a millimeter-wave radar based on frequency-modulated continuous wave. Filter and denoise real-time data; For the filtered and denoised real-time data, the corresponding point cloud dataset D is obtained according to the principle of frequency-modulated continuous wave.
3. The vehicle classification method according to claim 1, characterized in that, The prediction in step S2 using the extended Kalman filter algorithm includes: Linearize the state transition equation and the observation equation; Prediction is performed using the Kalman filter algorithm based on the linearized state transition equation and observation equation.
4. The vehicle classification method according to claim 1, characterized in that, The step S3 is followed by: A counting label is configured for each target vehicle. The initial value of the counting label is 0. When the target vehicle disappears in a certain frame of point cloud dataset D, the counting label is incremented by 1. When the number of count tags exceeds the preset tag threshold, the target vehicle is removed from the target vehicle set O. Once the target vehicle is re-matched in a point cloud dataset D of a certain frame, the count label is set to 0.
5. The vehicle classification method according to claim 1, characterized in that, Step S5 further includes: Traverse all frame point cloud datasets D and assign a car label to the first matched target vehicle. For target vehicles whose vehicle length is greater than or equal to a threshold value in x consecutive frames, they are locked as large vehicle labels, where x≥2; For target vehicles whose vehicle length is less than a threshold value in consecutive y frames, they are locked as car labels, where y≥2.
6. The vehicle classification method according to claim 1, characterized in that, The step S3 is followed by: Point cloud data exceeding the matching threshold are identified as noise and discarded.
7. A vehicle classification system, characterized in that, The system includes: The data acquisition module is used to acquire multiple frames of real-time data from the observation area and obtain the corresponding point cloud dataset D based on the real-time data, where D = {d1, d2, ..., d...} m }, d1~d m The data consists of point cloud data, each point cloud data includes velocity information v, distance information r, and angle information a; The data matching unit is used to match the point cloud data in the point cloud dataset D with the target vehicle data predicted by the extended Kalman filter algorithm, and add the target vehicles corresponding to multiple successfully matched point cloud data to the target vehicle set O; and to perform cluster analysis on the unmatched point cloud data using the DBSCAN algorithm, and add the target vehicles corresponding to multiple point cloud data exceeding the matching threshold to the target vehicle set O; and to perform two-dimensionalization on the point cloud data corresponding to all target vehicles to obtain the centroid position (x, y) and velocity (v) of each target vehicle. x v y Based on the constant acceleration model and the centroid position and velocity of the target vehicle, predict the centroid position and velocity of the target vehicle in the subsequent frame point cloud dataset D; compare whether a new target vehicle not present in the previous frame point cloud dataset D is matched in a certain frame point cloud dataset D; if a new target vehicle is matched, compare whether the centroid position and velocity of the new target vehicle are within the preset distance threshold and preset velocity threshold range of one of the predicted target vehicles; if so, delete the new target vehicle and map the point cloud data corresponding to the new target vehicle to the matched target vehicle; if not, add the new target vehicle to the target vehicle set O; if no new target vehicle is matched, keep the target vehicle set O unchanged. The length acquisition unit is used to obtain the distance between any two point cloud data based on the multiple point cloud data corresponding to each target vehicle in the target vehicle set O in the point cloud dataset D, and take the maximum distance as the vehicle length. The tag configuration unit is used to configure large and small vehicle tags for vehicles based on the vehicle length and threshold value. For target vehicles whose length is greater than or equal to the threshold value, a large vehicle tag is configured, and for target vehicles whose length is less than the threshold value, a small vehicle tag is configured.