A data fusion method and device

Abstract

The present invention relates to the field of radar technology, and in particular to a data fusion method and a device thereof. The method comprises the following steps: S1, using a lane calibration automatic division method to complete lane division and obtain a lane division result; S2, according to the lane division result, using a split target fusion method to fuse collected vehicle target data to obtain fused vehicle target information, complete the statistics of vehicle flow and vehicle speed, and effectively solve the problem of large truck target splitting and improve the accuracy of vehicle counting.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a data fusion method and a device thereof. Background Art

[0002] With the development of the times, the number of vehicles on the road has increased, and the collection of traffic vehicle data information has become an important part of modern urban management. It is of irreplaceable importance for maintaining smooth traffic, ensuring driving safety, optimizing urban planning and promoting sustainable development. Through accurate data collection, it is possible to monitor traffic flow in real time, promptly discover and solve congestion problems, and reduce the incidence of traffic accidents. With the rapid development of intelligent transportation and autonomous driving technology, accurate collection of vehicle data is crucial.

[0003] Radar technology is an advanced method of detecting the position of a target using electromagnetic waves. It can accurately detect the position, speed and direction of a target by emitting electromagnetic waves and receiving reflected echoes. In intelligent transportation systems, radar technology plays a vital role. It can not only monitor traffic flow and vehicle speed, but also improve transportation efficiency.

[0004] Existing technologies and devices can collect traffic data, but there are still many problems. One is that the installation process of the data collection device is complicated and difficult to achieve fast and accurate installation. Second, when collecting traffic data, manual lane calibration is time-consuming and has poor accuracy. Third, the radar collection of vehicles will cause vehicle target splitting due to complex reflection conditions and multipath effects, affecting the accuracy of data collection such as traffic flow and speed. Summary of the invention

[0005] The purpose of the present invention is to improve the accuracy of traffic data collection through a data fusion method; to improve the lane division accuracy and shorten the lane division time through automatic lane division; and to use an automatic fast calibration method to achieve fast and accurate installation of the device.

[0006] A data fusion method comprises the following steps:

[0007] S1, using the lane calibration automatic division method to complete the lane division and obtain the lane division result;

[0008] S2, according to the lane division result, using the split target fusion method, the collected vehicle target data is fused to obtain fused vehicle target information, and complete the statistics of vehicle flow and vehicle speed;

[0009] The collected target data includes lateral distance, lateral speed, lateral acceleration, longitudinal distance, longitudinal speed and longitudinal acceleration, wherein the longitudinal direction is the normal direction of the radar sensor collecting the data, and the lateral direction is the direction perpendicular to the normal direction of the radar sensor.

[0010] Preferably, S2 specifically includes the following steps:

[0011] S21: filtering out stationary targets in the collected vehicle target data to obtain filtered vehicle target data;

[0012] S22: grouping the filtered vehicle target data according to the lane division result, and sorting each group of filtered vehicles according to the target position using the TimSort algorithm to obtain sorted groups;

[0013] S23: In each of the sorted groups, using a split target fusion method, performing fusion judgment on adjacent targets to obtain target data to be fused;

[0014] S24: determining whether the number of frames in which the target to be fused appears continuously in the target data to be fused is greater than F frames, thereby reducing the interference of false targets and obtaining the filtered target data to be fused;

[0015] S25: Obtaining fused vehicle target information according to the screened target data to be fused;

[0016] S26: Performing statistics on the collected data according to the fused vehicle target information.

[0017] Preferably, step S23 specifically includes the following steps:

[0018] S231, for the sorted groups, set each target data as P j =[x1, x2, x3, y1, y2, y3], get the original data P, where j is the target number, x1 is the lateral distance, x2 is the lateral speed, x3 is the lateral acceleration, y1 is the longitudinal distance, y2 is the longitudinal speed, and y3 is the longitudinal acceleration;

[0019] S232, calculating the covariance matrix Cov(X1, X2, X3, Y1, Y2, Y3) according to the normalized data frame;

[0020] S232, calculating the eigenvalues ​​of the covariance matrix, sorting the eigenvalues ​​of the covariance matrix from large to small and calculating the corresponding contribution rates, summing the contribution rates to obtain a cumulative contribution rate, and selecting eigenvectors with cumulative contribution rates greater than or equal to 90% to form an eigenvector matrix A;

[0021] S234, converting the original data P into the eigenvector matrix A, completing dimensionality reduction to obtain processed data B = AP;

[0022] S235, select two adjacent targets in the data B after dimensionality reduction, and their contribution rates of k dimensions are , calculate the cumulative difference value f of the two targets;

[0023] S236, judging the target in the dimension-reduced data B according to the cumulative difference value f, if it is greater than 0.1, it is considered that it does not meet the fusion condition, and if it is not greater than 0.1, it is considered that it meets the fusion condition, and the target to be fused is obtained.

[0024] Preferably, the cumulative difference value f between the two targets is calculated using the following formula:

[0025] ;

[0026] In the formula Represents the r-th dimension data of the i-th target, Represents the r-th dimension data of the adjacent i+1-th target, l r is the contribution rate of the rth dimension, and k is the total number of dimensions.

[0027] Preferably, before step S1, a step of using a posture correction method to correct the posture of the data fusion device that collects vehicle target data is also included, specifically including:

[0028] S31: Determine whether the roll angle and pitch angle of the data fusion device are within the threshold [-5°, 5°], if not, prompt calibration, otherwise execute the next step;

[0029] S32: Use the DBSCAN-LR algorithm to determine whether the heading angle of the data fusion device is within a threshold value [-5°, 5°]. If it is within the threshold value, the data fusion device is facing the lane.

[0030] Preferably, step S1 specifically includes the following steps:

[0031] S11, calculating a thermal distribution value according to the lateral distance and the longitudinal distance, and drawing a distribution thermal map of the vehicle according to the thermal distribution value;

[0032] S12, using a multiple of the mean value of the thermal distribution value as a calibration threshold to offset the influence of noise;

[0033] S13, dividing the longitudinal distance into sections every 30 meters, obtaining an xf diagram at each section distance, and performing low-pass filtering;

[0034] S14, removing redundancy and completing the transverse points of all sections to obtain the final central axis point set;

[0035] S15, generating final lane line coordinates according to the center axis point set and lane width.

[0036] Preferably, the lane width is calculated as follows:

[0037] ;

[0038] Where N is the mode of the total number of points at the horizontal distance from the lane centerline of the section, x is the mean of the range of transverse distance points from the lane centerline of each section.

[0039] Preferably, in step S14, the redundancy removal process includes: points whose lateral distance from the center axis of the lane of each section is greatly different from the width of the conventional road are regarded as redundant points and are directly deleted; the completion process includes: finding the section and lane numbers that need to be completed, and completing the missing points by linear interpolation.

[0040] Based on the same concept, a data fusion device is also proposed, including a data acquisition unit, a data processing unit and a data storage unit;

[0041] The data acquisition unit is used to acquire lane division data and vehicle target data required for data fusion;

[0042] The data processing unit uses a lane calibration automatic division method to complete lane division and obtain a lane division result, and according to the lane division result, uses a split target fusion method to fuse the collected vehicle target data to obtain fused vehicle target information and complete the statistics of vehicle flow and vehicle speed;

[0043] The collected target data includes lateral distance, lateral velocity, lateral acceleration, longitudinal distance, longitudinal velocity and longitudinal acceleration, wherein the longitudinal direction is the normal direction of the radar sensor collecting the data, and the lateral direction is the direction perpendicular to the normal direction of the radar sensor;

[0044] The data storage unit is used to store data.

[0045] Preferably, the data acquisition unit includes a radar sensor and a visual sensor.

[0046] The radar sensor is used to collect lateral distance, lateral speed, lateral acceleration, longitudinal distance, longitudinal speed and longitudinal acceleration;

[0047] The visual sensor is used to synchronously collect video data for comparison with radar data.

[0048] Compared with the prior art, the present invention has the following beneficial effects: a data fusion method provided by the present invention uses a split target fusion method to sort the split vehicle targets and then fuse them, and after fusion, the vehicle information is counted, which can improve the accuracy of vehicle counting; an automatic lane division method can improve the lane division accuracy and shorten the lane division time; an automatic fast calibration and installation method can realize fast and accurate installation of the device. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of a data fusion method;

[0050] Figure 2 A flowchart for fusing the collected target data;

[0051] Figure 3 A schematic diagram of the roll angle, yaw angle and pitch angle of the radar sensor of the data fusion device in a three-axis coordinate system;

[0052] Figure 4 is a flow chart of the posture correction method;

[0053] Figure 5 A flow chart of the automatic lane division method in the lane demarcation phase;

[0054] Figure 6 is the denoised heat map;

[0055] Figure 7 (a) is a graph showing the lateral distance at the 30m intersection setting and the number of vehicles distributed at that distance;

[0056] Figure 7 (b) is a graph showing the lateral distance at the 83m intersection setting and the number of vehicles distributed at that distance;

[0057] Figure 7 (c) is a graph showing the lateral distance at the 136m intersection setting and the number of vehicles distributed at that distance;

[0058] Figure 7 (d) is a graph showing the lateral distance at the 189m intersection setting and the number of vehicles distributed at that distance;

[0059] Figure 8 (a) shows the lateral distance at the 30m intersection setting after filtering and the number of vehicles distributed at this distance;

[0060] Figure 8 (b) is the filtered lateral distance at the 83m intersection setting and the number of vehicles distributed at this distance;

[0061] Figure 8 (c) shows the lateral distance at the 136m intersection setting after filtering and the number of vehicles distributed at this distance;

[0062] Figure 8 (d) is the filtered lateral distance at the 189m intersection setting and the number of vehicles distributed at this distance;

[0063] Fig. 9 This is an example of the heading angle γ with the lane when the data fusion device is set up;

[0064] Fig.10 This is an external structural diagram of a data fusion device.

[0065] Markings in the figure: 1- radar sensor, 12- radar main beam symmetry axis of the radar sensor, 2- radar sensor, 3- data fusion device housing, 4- visual sensor, 5- angle sensor. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0067] Unless otherwise specified, in the description of the specific embodiments of the present invention, the terms indicating the orientation or position relationship such as "up", "down", "left", "right", "center", "inside", "outside", etc. are all expressions based on the orientation or position relationship described in the accompanying drawings, or are the orientation or position relationship of the invented product / data fusion device / device when it is usually used. These terms of orientation or position relationship are only for the convenience of describing the scheme of the present invention or simplifying the description in the specific embodiments, so as to facilitate the technicians to quickly understand the scheme, and do not indicate or imply that a specific device / component / element must have a specific orientation, or be constructed and operated in a specific position relationship, and therefore cannot be understood as a limitation of the present invention.

[0068] In addition, if the terms "horizontal", "vertical", "overhanging", "parallel" and the like appear, it does not mean that the corresponding devices / components / elements are required to be absolutely horizontal or vertical or overhanging or parallel, but may be slightly tilted or have deviations. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but may be slightly tilted. Alternatively, it can be simplified to mean that the corresponding devices / components / elements are set in directions such as "horizontal", "vertical", "overhanging", "parallel", etc., and can have an error / deviation of ±10% relative to the corresponding direction setting, more preferably an error / deviation within ±8%, more preferably an error / deviation within ±6%, more preferably an error / deviation within ±5%, and more preferably an error / deviation within ±4%. As long as the corresponding device / component / element is within the error / deviation range, it can still achieve its role in the scheme of the present invention.

[0069] In addition, the expressions “first”, “second”, “third”, etc., which appear in the terms, are merely used to distinguish the description of the same or similar components and should not be understood as emphasizing or implying the relative importance of specific components.

[0070] In addition, in the description of the embodiments of the present invention, "several", "plurality" and "a number" represent at least 2. It can be any number such as 2, 3, 4, 5, 6, 7, 8, 9, and even more than 9.

[0071] In addition, in the description of the technical solution of the present invention, unless otherwise clearly specified / defined / restricted, the terms "set", "install", "connect", "connected", "provided with", "laid", and "arranged" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection, and can be welding, riveting, bolting, threading, and other commonly used connection means in the field. This connection can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection, or an indirect connection through an intermediate medium, and it can be the internal connection of two elements.

[0072] Example 1

[0073] A data fusion method, the flow chart is as follows Figure 1 As shown, the following steps are included:

[0074] S1: Lane calibration phase, using the automatic lane division method in the lane calibration phase to complete lane division and obtain lane division results;

[0075] S2: Vehicle statistics stage, according to the lane division result, the split target fusion method is used to fuse the collected vehicle target data to obtain the fused vehicle target information, and complete the statistics of vehicle flow and vehicle speed. The collected target data includes lateral distance x1, lateral speed x2, lateral acceleration x3, longitudinal distance y1, longitudinal speed y2 and longitudinal acceleration y3. The lateral distance x1 is the distance of the vehicle target relative to the radar sensor of the data fusion device in the direction perpendicular to the radar normal, and the longitudinal distance y1 is the distance of the vehicle target relative to the normal direction of the radar sensor.

[0076] like Figure 2 As shown, fusing the collected target data includes the following steps:

[0077] S21: filtering out stationary targets in the collected vehicle target data to obtain filtered vehicle target data;

[0078] S22: grouping the filtered vehicle target data according to the lane division result, and sorting each group of the filtered vehicle target data according to the target position using the TimSort algorithm to obtain sorted groups;

[0079] S23: In each of the sorted groups, a split target fusion method is used to perform fusion judgment on adjacent targets to obtain target data to be fused;

[0080] S24: determining whether the motion duration of the target data to be fused is greater than F frames (preferably F=5), thereby reducing interference from false targets and obtaining the filtered target data to be fused;

[0081] S25: Obtaining fused vehicle target information according to the fusion of the target to be fused;

[0082] S26: Performing statistics on the collected data according to the fused vehicle target information.

[0083] In addition, there is an installation phase before the lane calibration phase, and the attitude correction method needs to be used to correct the attitude of the data fusion device that collects the target data;

[0084] The schematic diagram of the roll angle, yaw angle and pitch angle of the radar sensor of the data fusion device in the three-axis coordinate system is as follows: Figure 3 As shown, Figure 3 As shown, the center is the coordinate origin O, and a three-axis coordinate system is established. One of the axes is established with the geomagnetic north direction N as the reference, and another axis is established in the other direction X perpendicular to the N direction on the horizontal plane, and the third axis is perpendicular to the downward direction G of the road. Then planes abcd and efgh are established. When the radar sensor 1 adjusts the pitch angle, the plane where the radar main beam symmetry axis 12 is scanned is plane abcd, and plane efgh is a horizontal plane at the bottom of the radar sensor 1, which is located below plane NOX. There is an angle α between the bottom surface of the radar sensor 1 and plane efgh, and there is an angle θ between plane abcd and plane NOG. The angle θ is the angle between the direction of the symmetry axis 12 of the radar main beam and the geomagnetic north direction N. The radar position of the radar sensor 1 is taken as point O´, and the O´F direction perpendicular to the road surface downward is established. O´F is parallel to OG, and there is an angle V2 between the symmetry axis 12 of the radar main beam and O´F; the angle acquisition sensor 2 is used The current angle data of the radar sensor 1 is collected, including three current angle data: a roll angle, a yaw angle, and a pitch angle. The roll angle is the angle between the radar sensor 1 itself and the horizontal plane, that is, the angle at which the radar sensor 1 swings up and down relative to the horizontal plane. When swinging, the front of the radar sensor 1 faces the area to be measured on the highway, and its left and right sides swing up and down at a small angle. The angle is α. The yaw angle is the angle at which the working direction of the radar sensor 1 deviates from the magnetic north direction, that is, the angle between the direction of the symmetry axis 12 of the radar main beam of the radar sensor 1 and the magnetic north direction. The angle is θ. The pitch angle is the angle between the working direction of the radar sensor 1 and the downward direction perpendicular to the road surface, that is, the angle between the direction of the symmetry axis 12 of the radar main beam of the radar sensor 1 and the downward direction perpendicular to the road surface. The angle is V2. The reference direction of the roll angle is the horizontal direction, the reference direction of the yaw angle is the magnetic north direction, and the reference direction of the pitch angle is perpendicular to the road surface and downward.

[0085] The attitude correction method is to calibrate the heading angle, pitch angle and roll angle of the data fusion device. Use the attitude sensor to calibrate the pitch angle of the data fusion device and roll angle , rotating the data fusion device in the vertical direction and the left-right direction, when the offset value of the roll angle and the offset value of the pitch angle meet the set threshold value, the adjustment of the pitch angle and the roll angle is completed, and tightening the adjustment screw;

[0086] For the heading angle For calibration, the data fusion device is rotated in the horizontal direction (lane width direction), and the heading angle automatic calculation method is adopted to calculate the angle between the radar installation direction and the road centerline direction in real time. When it meets the set threshold, the adjustment screw is tightened.

[0087] Specifically, the posture correction method, the flow chart is as follows Figure 4 As shown, the following steps are included:

[0088] S31: Determine whether the roll angle and pitch angle of the data fusion device are within the threshold according to the attitude sensor If it is not within the threshold, a calibration prompt is displayed;

[0089] S32: After the offset value of the roll angle and the offset value of the pitch angle meet the threshold, vehicle target data of the radar data within a period of time is collected, and the Density-Based Spatial Clustering of Applications with Noise-Linear Regression algorithm, i.e., DBSCAN-LR algorithm, is used to determine whether the heading angle of the data fusion device is within the threshold. If it is within the threshold, the data fusion device is directed toward the lane.

[0090] Specifically, the automatic lane division method in step S1 is as follows: Figure 5 As shown, the following steps are included:

[0091] S11: The lateral distance and the longitudinal distance are divided at a certain interval to form a plurality of thermal frames of the same size. The value of the thermal frame is made equal to the total number of vehicles passing through the frame by a statistical method, and finally a distribution thermal map of all vehicles within a period of time within the detection range of the data fusion device is obtained. The size of the certain interval determines the precision of the thermal frame. A smaller interval is more conducive to lane division, but it will increase the amount of calculation. The reference value is 0.1-1 meter for the lateral distance and 1 meter for the longitudinal distance.

[0092] S12: The thermal distribution map may be contaminated by noise. It is necessary to use a certain calibration multiple of the mean of the thermal distribution value as a calibration threshold to perform mean processing to weaken the influence of noise and obtain a denoised thermal map. The denoised thermal map is as follows: Figure 6As shown, the calibration multiple is between 0.6 and 0.9, including 0.6 and 0.9; the noise points of the noise pollution have the characteristics of large values ​​and small numbers, so the noise dilution has been completed after the mean processing, reducing the influence of noise on the accuracy and performance of the thermal map;

[0093] S13: According to the denoised heat map, take the longitudinal distance as the reference direction, take (maximum longitudinal distance - minimum longitudinal distance) / 3 as the step size S in the longitudinal distance where data exists, and divide it into sections every step size S (such as 53 meters). At each section distance, the xf map is obtained, that is, the lateral distance - vehicle distribution number map at this distance is obtained. The lateral distance - vehicle distribution number map at this distance is shown in Figure 7. The horizontal axis represents the lateral distance, and the vertical axis represents the number of vehicles distributed at the lateral distance in the current section, which also represents the color depth in the heat map. Among them, Figure 7 (a) is the lateral distance - vehicle distribution number map at the 30m intersection setting; Figure 7 (b) is the lateral distance - vehicle distribution number map at the 83m intersection setting; Figure 7 (c) is the lateral distance - vehicle distribution number map at the 136m intersection setting; Figure 7 (d) is the lateral distance - vehicle distribution number map at the 189m intersection setting. Theoretically, the center axis of each lane in any time period is a statistical extreme point, that is, the xf diagram contains a pure waveform of several peaks of the lane. In fact, due to factors such as vehicle lane changes, the xf diagram shows the presence of high-frequency clutter. Therefore, a low-pass filtering method is used to perform low-pass filtering on the xf diagram of each section to obtain the filtered xf diagram of each section and the lateral distance point of high-frequency traffic flow of each section. The filtered xf diagram of each section is shown in the sub-figures of Figure 8. The horizontal axis represents the lateral distance, and the vertical axis represents the lateral distance in the current section. The statistical distribution of vehicles and the lateral distance points of high-frequency traffic flow in the cross section are shown in the "*" mark in Figure 8; Figure 8 (a) is the lateral distance at the 30m intersection setting after filtering - the number of vehicles distributed at this distance; Figure 8 (b) is the lateral distance at the 83m intersection setting after filtering - the number of vehicles distributed at this distance; Figure 8 (c) is the lateral distance at the 136m intersection setting after filtering - the number of vehicles distributed at this distance; Figure 8 (d) is the lateral distance at the 189m intersection setting after filtering - the number of vehicles distributed at this distance.

[0094] S14: According to the high-frequency traffic flow lateral distance points of each section, according to the expected total number of lanes N, the average of the range of high-frequency traffic flow lateral distance points of each section x calculates the lane width L, and the calculation formula for L is:

[0095] ;

[0096] The central axis point set generation method is used for the high-frequency traffic flow transverse points of all sections to obtain the final central axis point set;

[0097] S15: Generate final lane line coordinates based on the final center axis point set and the lane width L, offset the final center axis point set by L / 2, and then copy the leftmost or rightmost lane line by L to obtain N+1 lane lines, which are the lane line coordinates.

[0098] Specifically, in step S13, a low-pass filtering method is used to perform low-pass filtering on the xf graphs of each section to obtain filtered xf graphs of each section, which specifically includes the following steps:

[0099] S131: Design a low-pass filter. The required parameters are the cutoff frequency and the order. Consider a lane width wave in the theoretical xf wave as a reference cycle, and the cycle width is T W Since the frequency is inversely proportional to the period, filtering out high frequencies means filtering out periods less than T W If the horizontal distance is 0.2 meters accurate when dividing the thermal frame, the horizontal distance x is 60×1 / 0.2=300, and the sampling frequency is 300Hz for full sampling. The lowest frequency to be filtered is equivalent to the maximum width to be filtered, and the cutoff frequency is taken as the reference period width T W The corresponding frequency, since the main source of high-frequency clutter is that the trajectory of the traffic in the lane is divergent rather than strictly coincident with the central axis, this setting can make the clutter narrower than the lane width decay more quickly, and retain the outer lanes with lower traffic as much as possible. The cutoff frequency design here is not a fixed value. W The order can range from 2 to 4, and the xf wave of the xf graph is low-pass filtered according to the low-pass filter to obtain the filtered xf graph, that is, Figure 8;

[0100] S132: Peak-finding processing is performed on the filtered xf graph to obtain lateral distance points corresponding to the high-frequency traffic flow in each section. The lateral distance points are marked with asterisks in FIG8 . When selecting the peak-finding threshold, in order to reduce the impact caused by the traffic flow in the inner and outer lanes being lower than that in the middle lane, the maximum peak value is used as the reference value and multiplied by the calibration coefficient as the peak-finding threshold. The reference value of the calibration coefficient is between 0.2 and 0.5 to reduce lane missed detection, and finally the lateral distance points of the high-frequency traffic flow in each section are obtained.

[0101] Specifically, the central axis point set generation method in step S14 is used to obtain the final central axis point set, which is mainly performed based on the total number of lanes N for redundancy removal and completion, including the following steps:

[0102] Redundancy removal process: Judgment and processing are carried out based on the horizontal distance. The horizontal distance points of the high-frequency traffic flow of each cross-section are marked with "*" in Figure 8. At each cross-section, the spacing between adjacent distance points is calculated by forward difference as shown in Table 1 below. The correct difference value should be approximately equal to the highway standard width divided by the horizontal division accuracy of the heat map. In this embodiment, it should be approximately equal to 3.75 / 0.2 = 18.75. Points that are significantly narrower than the highway width and wider than the highway width are regarded as redundant points and directly removed. The removal rule is as follows: When the difference value is much smaller than the highway width value, the right adjacent point of the abnormal value position is removed. For example, at 189m, the difference value is 5 (corresponding to an actual width of 1m), and the abnormal position is 2. Therefore, the 3rd point 148 in the horizontal points is removed; when the difference value is much larger than the highway width value, according to the principle of deleting the previous point on the left and the subsequent point on the right, that is, if the difference abnormal position is greater than half of the number of difference values, it is on the right, and all points on the right of this point need to be deleted. For example, at 30m, the 4th difference value is 76 (corresponding to an actual width of 15.2m), and half of the number of difference values is 3. Therefore, the 5th point in the horizontal points and all points on its right need to be deleted. If the difference abnormal position is less than half of the number of difference values, it is on the left, and all points on the left of this point need to be deleted. For example, at 83m, the 1st difference value is 34 (corresponding to an actual width of 6.8m), and the 1st point in the horizontal points and all points on its left need to be deleted.

[0103] Table 1 Example table of calculating forward difference and difference abnormal position in the redundancy removal process

[0104]

[0105] Completion process: For cross-sections where the number of horizontal distance points of the high-frequency traffic flow is less than the lane mode N, completion is required. In the aforementioned attitude correction, the course angle has been ensured to be the minimum angle. The course angle γ between the data fusion device and the lane during erection is shown in Figure 8 for example. From the trigonometric relationship, the offset value , so that the abscissa at each point of the lane - the corresponding offset off can be approximately normal to the normal line (x = 0). Therefore, the completion rule is as follows: Calculate the lane number to be completed: Traverse each high-frequency traffic flow horizontal point on the cross-section y M (assuming there are a total of M points, M < N), subtract each point at the cross-section y N with a horizontal point number of N in turn, and then subtract the offset off to obtain N differences (the value is M i - N j - off). Record the position loc of the point with the smallest absolute value in the differences. Then, a total of M locs are obtained. The M locs represent the lane numbers that match successfully and do not need to be completed, and at the same time, the set of lane numbers to be completed is obtained; Calculate the coordinates of the completed points: Traverse the set of lane numbers to be completed. According to the above offset value calculation formula, record y = y N - y M , then y MThe coordinates of the zth lane completion point are ( ,y M ). After the above two steps, the lane number and the corresponding coordinate value to be completed in the section to be completed can be obtained.

[0106] S143: Repeat steps S141-S142 for the sections whose number of points in the point set of the lane centerline lateral distance points is not equal to the total number of lanes N, until the number of lane centerline lateral distance points obtained for each section is consistent, which is N. The obtained point set is the final centerline point set.

[0107] In this embodiment, the vehicle target completes the lane division of the vehicle target by analyzing the lane line coordinates obtained by automatic lane division and combining its own coordinate information. A specific division result is shown in Table 2. Table 2 is as follows:

[0108] Table 2 Data table of vehicle target after lane division

[0109]

[0110] Furthermore, the fusion judgment method of step S23 judges the fusion of split targets according to the position, speed and acceleration of the vehicle to obtain the target data to be fused, which specifically includes the following steps:

[0111] S231: For the sorted groups, each target is P j =[x1, x2, x3, y1, y2, y3] data, j is the target number, including lateral distance x1, lateral speed x2, lateral acceleration x3, longitudinal distance y1, longitudinal speed y2 and longitudinal acceleration y3, the data of each frame is P=[P1, P2, P3, ... Pn], n is the number of targets, in order to reduce the amount of calculation while ensuring that the data features are not lost as much as possible, it is necessary to reduce the dimension of the data. First, the Min-Max normalization method is used to process the current data frame to obtain the normalized data frame;

[0112] S232: Calculate the covariance matrix Cov(X1, X2, X3, Y1, Y2, Y3) according to the normalized data frame;

[0113] S233: Calculate the eigenvalues ​​and eigenvectors of the covariance matrix, then sort the eigenvalues ​​from large to small and calculate the corresponding contribution rates, select the eigenvectors whose cumulative contribution rates are greater than or equal to 90%, and define the total number as k, and

[0114] The eigenvectors are used as row vectors to form the eigenvector matrix A, and the contribution rate is calculated using the following formula:

[0115] ;

[0116] in, is the contribution rate of the qth eigenvalue, For the q The sum of multiple contribution rates is the cumulative contribution rate;

[0117] S234: Use the eigenvector matrix A to perform data dimensionality reduction on the original data P, and the reduced-dimensional data B = AP, wherein the reduced-dimensional data contains k dimensions ;

[0118] S235: Select two adjacent targets in the data B after dimensionality reduction, and the contribution rates of k dimensions are , calculate the cumulative difference value f of the two targets, which is calculated using the following formula:

[0119] ;

[0120] In the formula, Represents the r-th dimension data of the i-th target, Represents the r-th dimension data of the adjacent i+1-th target, l r is the contribution rate of the rth dimension, and k is the total number of dimensions.

[0121] S236: According to the cumulative difference value f, the target in the reduced-dimensional data B is judged. If it is greater than 0.1, it is considered that the fusion condition is not met. If it is not greater than 0.1, it is considered that the fusion condition is met. The target to be fused is obtained and saved in the table to be fused. Then, step S235 is repeated for all targets until all the targets are traversed and the repetition is stopped.

[0122] Table 3 A specific set of initial data collected

[0123]

[0124] Table 3 is a specific set of initial data collected, and one frame of data when the split target fusion method is not performed is shown in Table 4. Table 4 is as follows:

[0125] Table 4 Data of one frame when the split target fusion method is not performed

[0126]

[0127] Among them, a large vehicle is identified as two targets, namely, targets with vehicle IDs 21 and 20, which causes a deviation in vehicle counting. Steps S231-S236 are used to judge the split targets. Since the amount of data in one frame is smaller than the data dimension, and too little data may lead to poor dimensionality reduction performance, some historical data can be filled. The filled data is shown in Table 5:

[0128] Table 5 Filled data

[0129]

[0130] The initial matrix P composed of the longitudinal distance, longitudinal velocity, longitudinal acceleration, lateral distance, lateral velocity and lateral acceleration in the selected data is:

[0131]

[0132] After normalizing P, the eigenvalues ​​and eigenvectors of the defense-cooperation difference matrix are calculated. The calculation results are as follows:

[0133] Eigenvalues: ;

[0134] Eigenvectors:

[0135] The contribution rates corresponding to the calculated eigenvalues ​​are: [9.58*10 -1 , 4.18*10 -2 ,8.90*10 -4 , 3.03*10 -5 , 6.59*10 -6 , 2.85*10 -7 ], it can be concluded that the cumulative contribution rate of the first eigenvalue can meet the threshold, so the dimension reduction matrix A is:

[0136] ;

[0137] After that, the data after dimension reduction is calculated as follows:

[0138] ;

[0139] Among them, the first two data are real data, and the data after are filled historical data. When judging the fusion condition, only the first two need to be judged. Substituting into formula S235, the cumulative difference value is 0.043, which meets the fusion condition. The attribute of the target after fusion takes the average value of the target before fusion. The fused data is shown in Table 6:

[0140] Table 6 Fusion data

[0141]

[0142] The purpose of obtaining the target data to be fused through the fusion judgment method of S23 is to reduce the dimension of the data. The contribution value is calculated by the eigenvalue of the covariance matrix. The transformation matrix composed of eigenvectors is essentially a coordinate transformation. The larger the corresponding eigenvalue contribution value, the more overall information is contained in this dimension after the transformation. At the same time, calculating the cumulative difference after dimensionality reduction can also avoid the problem of threshold judgment for each dimension. The split targets of the radar have similar characteristics, so the fusion target is determined by judging the difference.

[0143] Example 2

[0144] A data fusion device comprises an angle sensor, a radar sensor, a visual sensor, a data processing unit, a traffic data investigation device housing, and a data storage unit. The angle sensor, the radar sensor, the visual sensor and the traffic data investigation device housing are combined to form the data fusion device. The external structure is as follows: Fig.10 Wherein 2 is an angle sensor, 3 is a radar sensor, 4 is a data fusion device housing, and 5 is a visual sensor. The angle sensor 2, the radar sensor 3, and the visual sensor 5 are all parallel to the bottom surface of the traffic data investigation device housing 4. The data processing unit and the data storage unit are inside the traffic data investigation device. The data collected by the angle sensor 2 and the radar sensor 3 are all passed through the data processing unit and stored in the data storage unit.

[0145] The angle sensor 2 collects roll angle and pitch angle data and transmits the data to the data processing unit.

[0146] The radar sensor 3 collects radar data and transmits it to the data processing unit.

[0147] The visual sensor 5 collects visual data and transmits it to the data processing unit.

[0148] The data processing unit transmits the processed data to the data storage unit.

[0149] The data processing unit stores all data.

[0150] The data processing unit and the data processing unit are inside the data fusion device. Fig.10 Not shown in .

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A data fusion method, characterized in that: The following steps are involved: S1, using the lane calibration automatic division method to complete the lane division and obtain the lane division result; S2, according to the lane division result, using the split target fusion method, the collected vehicle target data is fused to obtain fused vehicle target information, and complete the statistics of vehicle flow and vehicle speed; The collected target data includes lateral distance, lateral velocity, lateral acceleration, longitudinal distance, longitudinal velocity and longitudinal acceleration, wherein the longitudinal direction is the normal direction of the radar sensor collecting the data, and the lateral direction is the direction perpendicular to the normal direction of the radar sensor; In the process of fusing the collected vehicle target data using the split target fusion method, the collected target data are first grouped, and then each target data is set to obtain the original data P; based on the normalized data frame, the covariance matrix of the original data P is calculated; Calculate the eigenvalues ​​of the covariance matrix, sort the eigenvalues ​​of the covariance matrix from large to small and calculate the corresponding contribution rates, sum the contribution rates to obtain the cumulative contribution rates, and select eigenvectors whose cumulative contribution rates are greater than or equal to a preset contribution rate threshold to form an eigenvector matrix A; Use the eigenvector matrix A to reduce the dimension of the original data P, and the reduced-dimensional data B = AP; Select two adjacent targets in the dimension-reduced data B, and calculate the cumulative difference value f of the two targets; According to the cumulative difference value f, determining the target in the dimension-reduced data B; If it is not greater than the preset threshold, it is considered to meet the fusion conditions and the target to be fused is obtained.

2. A data fusion method according to claim 1, characterized in that: S2 specifically includes the following steps: S21: filtering out stationary targets in the collected vehicle target data to obtain filtered vehicle target data; S22: grouping the filtered vehicle target data according to the lane division result, and sorting each group of filtered vehicles according to the target position using the TimSort algorithm to obtain sorted groups; S23: In each of the sorted groups, using a split target fusion method, performing fusion judgment on adjacent targets to obtain target data to be fused; S24: determining whether the number of frames in which the target to be fused appears continuously in the target data to be fused is greater than F frames, thereby reducing the interference of false targets and obtaining the filtered target data to be fused; S25: Obtaining fused vehicle target information according to the screened target data to be fused; S26: Performing statistics on the collected data according to the fused vehicle target information.

3. A data fusion method according to claim 2, characterized in that: Step S23 specifically includes the following steps: S231, for the sorted groups, set each target data as P j =[x1, x2, x3, y1, y2, y3], get the original data P, where j is the target number, x1 is the lateral distance, x2 is the lateral speed, x3 is the lateral acceleration, y1 is the longitudinal distance, y2 is the longitudinal speed, and y3 is the longitudinal acceleration; S232, calculating the covariance matrix Cov(X1, X2, X3, Y1, Y2, Y3) according to the normalized data frame; S233, calculating the eigenvalues ​​of the covariance matrix, sorting the eigenvalues ​​of the covariance matrix from large to small and calculating the corresponding contribution rates, summing the contribution rates to obtain a cumulative contribution rate, and selecting eigenvectors with cumulative contribution rates greater than or equal to 90% to form an eigenvector matrix A; S234, using the eigenvector matrix A to perform data dimension reduction on the original data P, and the data after dimension reduction is B = AP; S235, select two adjacent targets in the data B after dimensionality reduction, and their contribution rates of k dimensions are , calculate the cumulative difference value f of the two targets; S236, judging the target in the dimension-reduced data B according to the cumulative difference value f, if it is greater than 0.1, it is considered that it does not meet the fusion condition, and if it is not greater than 0.1, it is considered that it meets the fusion condition, and the target to be fused is obtained.

4. A data fusion method according to claim 3, characterized in that: The cumulative difference value f of the two targets is calculated as follows: ； In the formula Represents the r-th dimension data of the i-th target, Represents the r-th dimension data of the adjacent i+1-th target, l r is the contribution rate of the rth dimension, and k is the total number of dimensions.

5. A data fusion method according to claim 1, characterized in that: Before step S1, the method further includes the step of using a posture correction method to correct the posture of the data fusion device for collecting vehicle target data, which specifically includes: S31: Determine whether the roll angle and pitch angle of the data fusion device are within the threshold [-5°, 5°], if not, prompt calibration, otherwise execute the next step; S32: Use the DBSCAN-LR algorithm to determine whether the heading angle of the data fusion device is within a threshold value [-5°, 5°]. If it is within the threshold value, the data fusion device is facing the lane.

6. A data fusion method according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11, calculating a thermal distribution value according to the lateral distance and the longitudinal distance, and drawing a distribution thermal map of the vehicle according to the thermal distribution value; S12, using a multiple of the mean value of the thermal distribution value as a calibration threshold to offset the influence of noise; S13, dividing the longitudinal distance into sections every 30 meters, obtaining an xf diagram at each section distance, and performing low-pass filtering; S14, removing redundancy and completing the transverse points of all sections to obtain the final central axis point set; S15, generating final lane line coordinates according to the center axis point set and lane width.

7. A data fusion method according to claim 6, characterized in that: The formula for calculating lane width is: ； Where N is the mode of the total number of points at the horizontal distance from the lane centerline of the section, x is the mean of the range of transverse distance points from the lane centerline of each section.

8. A data fusion method according to claim 6, characterized in that: In step S14, the redundancy removal process includes: the points whose lateral distance from the lane centerline of each section is significantly different from the width of the conventional highway are regarded as redundant points and are directly deleted; the completion process includes: finding the section and lane number that need to be completed, and completing the missing points by linear interpolation.

9. A data fusion device, characterized in that: It includes a data acquisition unit, a data processing unit and a data storage unit; The data acquisition unit is used to acquire lane division data and vehicle target data required for data fusion; The data processing unit uses a lane calibration automatic division method to complete lane division and obtain a lane division result, and according to the lane division result, uses a split target fusion method to fuse the collected vehicle target data to obtain fused vehicle target information and complete the statistics of vehicle flow and vehicle speed; The collected target data includes lateral distance, lateral velocity, lateral acceleration, longitudinal distance, longitudinal velocity and longitudinal acceleration, wherein the longitudinal direction is the normal direction of the radar sensor collecting the data, and the lateral direction is the direction perpendicular to the normal direction of the radar sensor; The data storage unit is used to store data; In the process of fusing the collected vehicle target data using the split target fusion method, the collected target data are first grouped, and then each target data is set to obtain the original data P; based on the normalized data frame, the covariance matrix of the original data P is calculated; Calculate the eigenvalues ​​of the covariance matrix, sort the eigenvalues ​​of the covariance matrix from large to small and calculate the corresponding contribution rates, sum the contribution rates to obtain the cumulative contribution rates, and select eigenvectors whose cumulative contribution rates are greater than or equal to a preset contribution rate threshold to form an eigenvector matrix A; Use the eigenvector matrix A to reduce the dimension of the original data P, and the reduced-dimensional data B = AP; Select two adjacent targets in the dimension-reduced data B, and calculate the cumulative difference value f of the two targets; According to the cumulative difference value f, determining the target in the dimension-reduced data B; If it is not greater than the preset threshold, it is considered to meet the fusion conditions and the target to be fused is obtained.

10. A data fusion device according to claim 9, characterized in that: The data acquisition unit includes a radar sensor and a visual sensor. The radar sensor is used to collect lateral distance, lateral speed, lateral acceleration, longitudinal distance, longitudinal speed and longitudinal acceleration; The visual sensor is used to synchronously collect video data for comparison with radar data.

Images