A method for fusing back-end data of ranging devices based on multi-kinematic model fusion and its application

Abstract

This invention discloses a data fusion method for the backend of a ranging device based on multi-kinematic model fusion, comprising the following steps: Step 1, treating the motion of the target object measured by two ranging devices as a weighted fusion of four kinematic models. Step 2, predicting the trajectory based on the weighted fused kinematic model of the target object at time k-1. Step 3, selecting the measurement value of the two ranging devices at time k with the predicted value at time k-1 that has the closest Euclidean distance as the reliable measurement value. Step 4, updating the weights of each sub-model by calculating the prediction error of each sub-model individually at time k-1. Step 5, multiplying the temporary weights by the Markov state transition matrix to obtain the final corrected model weights at time k. Step 6, correcting the Markov state transition matrix based on all updated sub-model weights at time k and the probability weights of each sub-model at time k-1. This invention can obtain reliable and continuous target distance measurement results.

Description

Technical Field

[0001] This invention relates to the field of ranging technology, and in particular to a data fusion method for the backend of ranging equipment based on multi-kinematic model fusion, which can be used for target-level fusion of millimeter-wave radar and lidar, lidar and machine vision camera, or millimeter-wave radar and machine vision camera. Background Technology

[0002] Accurate and reliable obstacle location information is a crucial task in autonomous driving, as obstacle information constantly influences the autonomous vehicle's decision-making, trajectory planning, and other behaviors. Millimeter-wave radar, with its advantages of high range accuracy and strong environmental adaptability, is currently equipped in most unmanned aerial vehicles. Visual sensors, capable of accurately classifying obstacles and recognizing signs and traffic lights, are equally indispensable. However, both types of sensors face the challenges of measurement noise and false detections under various interference conditions. For millimeter-wave radar, electromagnetic interference can lead to false alarms; for visual sensors, incorrect clustering and the limited accuracy of the training model can cause target loss, jumps, and other problems. To compensate for false detections and improve measurement accuracy, the fusion of the two sensors is essential.

[0003] Various forms of sensor fusion methods can be divided into three categories: high semantic level fusion, which is considered a track fusion method when each target has been detected and tracked. The literature (ROChavez-Garcia and O. Aycard, "Multiple Sensor Fusion and Classification for Moving Object Detection and Tracking," in IEEE Transactions on Intelligent Transportation Systems, vol.17, no.2, pp.525-534, Feb.2016.) proposes a target tracking method that fuses detection level and trajectory level into a target list and a trajectory list. It develops a general fusion framework based on Dempster-Shafer theory, which uses cameras for target classification and lidar / millimeter-wave radar for detection. The literature (FJBotha, CEvan Daalen and J. Treurnicht, "Data fusion of radar and stereo vision for detection and tracking of moving objects," 2016 Pattern Recognition Association of South Africa and Robotics and Mechatronics International Conference (PRASA-RobMech), 2016, pp. 1-7.) obtains the target's motion through hierarchical clustering based on the image feature trajectory generated by radar state estimation. The literature (Kim, DY, and M. Jeon (2014), "Data Fusion of Radar and Image Measurements for Multi-object Tracking via Kalman Filtering," Information Sciences 278, 641-652.) utilizes a CCD camera to compensate for the low angular resolution of the radar.The paper (S. Kim, H. Kim, W. Yoo and K. Huh, "Sensor Fusion Algorithm Design in Detecting Vehicles Using Laser Scanner and Stereo Vision," in IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 4, pp. 1072-1084, April 2016.) predicts scanner data every 1 ms using Kalman filtering and matches it with stereo vision. The paper (T. Vu, O. Aycard and F. Tango, "Object perception for intelligent vehicle applications: A multi-sensor fusion approach," 2014 IEEE Intelligent Vehicles Symposium Proceedings, 2014, pp. 774-780.) develops a frontal target perception module to analyze inputs from multiple sensors and perform tasks such as target tracking and classification. The paper (X.Wang, L.Xu, H.Sun, J.Xin and N.Zheng, "On-Road Vehicle Detection and Tracking Using MMW Radar and Monovision Fusion," in IEEE Transactions on Intelligent Transportation Systems, vol.17, no.7, pp.2075-2084, July 2016.) proposes a visual detection method that utilizes the region of interest (ROI) provided by millimeter-wave radar to limit radar false alarms.The paper (K. Lee, Y. Kanzawa, M. Derry and MR. James, "Multi-Target Track-to-Track Fusion Based on Permutation Matrix Track Association," 2018 IEEE Intelligent Vehicles Symposium (IV), 2018, pp. 465-470, doi:10.1109 / IVS.2018.8500433.) developed a Permutation Matrix Track Association (PMTA) algorithm, which independently processes measurement data from multiple sensors, creates a trajectory of an object, and then fuses the list of trajectories of the object by filtering.

[0004] Besides high-level semantic fusion, low-level semantic fusion aims to process raw data from radar and cameras, typically containing pixels and clusters. Hybrid fusion structures consist of both low-level and high-level fusion. The paper (A. Milella, G. Reina, J. Underwood and B. Douillard, "Combining radar and vision for self-supervised ground segmentation in outdoor environments," 2011 IEEE / RSJ International Conference on Intelligent Robots and Systems, 2011, pp. 255-260.) proposes a self-supervised classification system incorporating radar and vision sensors, selecting ground patches from camera images through radar measurements. The paper (S. Wu, S. Decker, P. Chang, T. Camus and J. Eledath, "Collision Sensing by StereoVision and Radar Sensor Fusion," in IEEE Transactions on Intelligent Transportation Systems, vol. 10, no. 4, pp. 606-614, Dec. 2009.) creates a fused contour by fusing radar observations and visual contours, and then tracks the fused contour through rigid body constraints. The literature (W. Huang, Z. Zhang, W. Li, and J. Tian, ​​“Moving object tracking based on millimeter-wave radar and vision sensor,” J. Appl. Sci. Eng., vol. 21, no. 4, pp. 609–614, 2018.) detects candidate regions in images using adaptive background subtraction and identifies radar targets within these regions. While the hybrid structure is flexible, it is costly and computationally intensive.

[0005] The aforementioned work employs various methods such as evidence theory, filtering, track fusion, clustering, and region of interest (ROI) segmentation to improve the tracking performance of millimeter-wave radar / machine vision cameras. However, when applied to autonomous vehicle environments, these methods do not consider the motion characteristics and state of the target object. For example, the motion state of the target vehicle can be described using a kinematic model. Currently, there is a lack of publicly available data on the discrimination and fusion of measurement signals based on the kinematic physical characteristics of the tracked target. Summary of the Invention

[0006] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method for fusing back-end data of ranging devices based on the fusion of multiple kinematic models.

[0007] Another objective of this invention is to provide the application of the distance measuring device backend data fusion method in intelligent driving.

[0008] The technical solution adopted to achieve the purpose of this invention is:

[0009] A method for fusing backend data of a ranging device based on multi-kinematic model fusion includes the following steps:

[0010] Step 1: After unifying the coordinates of the two ranging devices, the motion of the target object measured by the two devices is regarded as a weighted fusion of four kinematic models. The four kinematic models are regarded as four sub-models, namely, constant velocity model, constant acceleration model, constant rotation and velocity model, and constant rotation and acceleration model.

[0011] Step 2, based on the weighted fusion kinematic model of the target object at time k-1... The initial value is used to predict the state vector at time k.

[0012] Step 3: At time k, measure the current state vector X of the target object using a ranging device. k The current state vectors measured by the two ranging devices are denoted as follows: and The two results were combined with the kinematic model predictions at time k-1. For comparison, the measurement results were screened, and the results were selected. and Zhongyu The nearest Euclidean distance is taken as the reliable measurement value, and the reliable measurement value is output in this step.

[0013] Step 4, with As input, calculate the prediction value of each sub-model at time k-1 individually. by Reliable measurement at time k The distance is used to update the temporary weights of each sub-model. And normalize it;

[0014] Step 5, set temporary weights With Markov state transition matrix P k-1 Multiplying them yields the final corrected sub-model weights at time k. And normalize it;

[0015] Step 6: Based on the updated sub-model weights at time k With each sub-model at time k-1 Weight For the Markov state transition matrix P k-1 After correction, P is obtained. k It will be used at time k+1.

[0016] In the above technical solution, the two ranging devices in step 1 are millimeter-wave radar and lidar, lidar and machine vision camera, or millimeter-wave radar and machine vision camera.

[0017] In the above technical solution, the constant velocity model in step 1 is:

[0018] CV k =[x k y k v k θ k 0 0] T (1)

[0019]

[0020] The constant acceleration model is as follows:

[0021] CA k =[x k y k v k θ k a k 0] T (3)

[0022]

[0023] The constant rotation rate and speed model is as follows:

[0024] CTRV k =[x k y k v k θ k 0 ω k ] T (5)

[0025]

[0026] The constant rotation rate and acceleration model is as follows:

[0027] CTRA k =[x k y k v k θ k ak ω k ] T (7)

[0028]

[0029] In formula (8):

[0030]

[0031]

[0032] Among them, CV k CA k CTRV k CTRA k These are the state vectors described by the constant velocity model, constant acceleration model, constant speed and velocity model, and constant speed and acceleration model, respectively, where x k y k v k θ k a k ω k These represent the target's x-coordinate, y-coordinate, heading angle, acceleration, and angular velocity in UTM coordinates at time k, respectively. k To calculate the step size.

[0033] In the above technical solution, the state vector X k =[x kkkkkk Each represents the x-coordinate, y-coordinate, velocity, heading angle, acceleration, and angular velocity of the detected target at time k in the global coordinate system.

[0034] The predicted values ​​of the weighted fusion kinematics model The result can be obtained using formula (9):

[0035]

[0036] This represents the predicted state at time k+1 calculated by the weighted fusion of four kinematic models at time k. to These are the weights for each sub-model.

[0037] In the above technical solution, in step 4, the temporary weights of the sub-model... The update process is shown in formulas (10), (11), (12), and (13):

[0038]

[0039]

[0040]

[0041]

[0042] For the target x and y coordinates predicted by the fusion model, The x and y coordinates of the target measured by the ranging device. Similarly, 'a' is an index indicating the distance between the azimuth distances measured by the two ranging devices and the predicted values ​​by the fusion kinematic model. Formula (11) reflects the iterative update process of the temporary weights of each sub-model. Formula (11) normalizes this process, and the sum of the weights of each sub-model should equal 1. This is the updated, unnormalized temporary weight vector. This is the normalized temporary weight.

[0043] In the above technical solution, in step 5, the sub-model weights The calculation process is as shown in formula (14):

[0044]

[0045]

[0046] Where μ k P represents the final sub-model weights after Markov state transition matrix operations and normalization. k This is the Markov state transition matrix;

[0047]

[0048] In the above technical solution, in step 6, the Markov state transition matrix is ​​corrected according to the current and historical weight coefficients of each sub-model, as shown in formulas (15) and (16).

[0049]

[0050]

[0051] in Let i be the probability of model i at the current time. Let be the probability of model i at the previous time step. Let be the element in the i-th row and j-th column of the Markov state transition matrix.

[0052] In the above technical solution, the distance measuring device backend data fusion method based on multi-kinematic model fusion further includes step 8, wherein when neither of the two distance measuring devices has a false detection, the two distance measuring devices are weighted and fused to smooth their output.

[0053] In the above technical solution, the weighted fusion process is as follows:

[0054]

[0055]

[0056]

[0057] k x ,k y The weights are assigned to the two ranging devices in the xy direction.

[0058] Another aspect of the present invention is the application of the multi-kinematic model fusion ranging device backend data fusion method in intelligent driving.

[0059] In the above technical solution, the dangerous vehicle ahead is tracked as a target object, and the reliable measurement value is used. or Input to the intelligent driving control system.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] 1. This invention fuses backend data from two ranging devices, which can be millimeter-wave radar and lidar, lidar and machine vision camera, or millimeter-wave radar and machine vision camera. This invention performs reliability calculation on the outputs of the two ranging devices, intelligently discarding distance misdetection, false detection, and tracking loss caused by external interference, and obtaining reliable and continuous target distance measurement results.

[0062] 2. This invention uses the predicted value from the previous moment as a metric to intelligently identify the problem of output value jumps when a sensor generates false detections. It can discard such jumps and use another sensor for complementation. To ensure prediction accuracy, each kinematic model is predicted independently. The composition of each component in the fused kinematic model is determined based on the error of each independent kinematic prediction. The model is iteratively updated continuously based on the prediction error to ensure prediction accuracy. Attached Figure Description

[0063] Figure 1 This is caused by false detections from the ranging device;

[0064] Figure 2 This is the overall architecture of the algorithm of this invention;

[0065] Figure 3 It is an output smoothing method;

[0066] Figure 4 It compares the prediction errors of each model;

[0067] Figure 5 It is the result of measurement selection;

[0068] Figure 6 It is a measurement error caused by data fusion;

[0069] Figure 7 It is the distribution of measurement errors in data fusion. Detailed Implementation

[0070] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0071] Example 1

[0072] The two ranging devices in this embodiment are a millimeter-wave radar and a machine vision camera.

[0073] With the millimeter-wave radar and machine vision camera fixed, the invention uses the millimeter-wave radar and machine vision camera to track the target in front, and performs stable single-target tracking.

[0074] A method for fusing backend data of a ranging device based on multi-kinematic model fusion includes the following steps:

[0075] Step 1: After unifying the coordinates of the machine vision camera and the millimeter-wave radar, the motion of the target object measured by the millimeter-wave radar and the machine vision camera is regarded as a weighted fusion of four kinematic models (four sub-models) in this coordinate system, namely the constant velocity model (CV), the constant acceleration model (CA), the constant rate of rotation and velocity model (CTRV), and the constant rate of rotation and acceleration model (CTRA).

[0076] The constant velocity model (CV) is as follows:

[0077] CV k =[x k y k v k θ k 0 0] T (1)

[0078]

[0079] The constant acceleration model (CA) is as follows:

[0080] CA k =[x k y k v k θ k a k 0]T (3)

[0081]

[0082] The constant speed and velocity model (CTRV) is as follows:

[0083] CTRV k =[x k y k v k θ k 0 ω k ] T (5)

[0084]

[0085] The constant rotation and acceleration model (CTRA) is as follows:

[0086] CTRA k =[x k y k v k θ k a k ω k ] T (7)

[0087]

[0088] In formula (8):

[0089]

[0090]

[0091] Among them, CV k CA k CTRV k CTRA k These are the state vectors described by the constant velocity model, constant acceleration model, constant speed and velocity model, and constant speed and acceleration model, respectively, where x k y k v k θ k a k ω k These represent the target's x-coordinate, y-coordinate, heading angle, acceleration, and angular velocity in UTM coordinates at time k, respectively. k To calculate the step size.

[0092] Step 2. Based on the weighted fusion kinematic model of the target object at time k-1, The initial value is obtained by predicting its state vector at time k (including coordinates, heading angle, acceleration, and angular velocity in the global environment). The motion of the target object is considered as a weighted fusion of four kinematic models. This represents the predicted future state of the target object at time k based on the fused kinematic model at time k-1. For example... Figure 4 As shown, the weighted fusion kinematics model has the lowest prediction error.

[0093] State vector X k =[x kkkkkk Each represents the x-coordinate, y-coordinate, velocity, heading angle, acceleration, and angular velocity of the detected target at time k in the global coordinate system.

[0094] The predicted values ​​of the weighted fusion kinematics model The result can be obtained using formula (9):

[0095]

[0096] This represents the predicted state at time k+1 calculated by the weighted fusion of four kinematic models at time k. to These are the weights for each sub-model.

[0097] Step 3: At time k, measure the current state vector X of the target object using a ranging device. k The current state vectors measured by the two ranging devices are denoted as follows: and The two results were combined with the kinematic model predictions at time k-1. For comparison, the measurement results were screened, and the results were selected. and Zhongyu The nearest Euclidean distance is taken as the reliable measurement value, and the reliable measurement value is output in this step. If a false detection occurs at this moment, the false detection value will deviate from the predicted value of the fused kinematics model, and the measurement from another ranging device will be judged by the system as the reliable measurement value at that moment. For example... Figure 5 As shown.

[0098] Step 4, with As input, calculate the prediction value of each sub-model at time k-1 individually. by Reliable measurement at time k The distance is used to update the temporary weights of each sub-model. Then normalize; the correction coefficient of each sub-model weight is inversely proportional to its prediction error at the previous time step. After correcting the weight of each sub-model, normalize the weights of each sub-model so that their sum is 1 to obtain the temporary weight.

[0099] Temporary weights of the sub-model The update process is shown in formulas (10), (11), (12), and (13):

[0100]

[0101]

[0102]

[0103]

[0104] For the target x and y coordinates predicted by the fusion model, The x and y coordinates of the target measured by the ranging device. Similarly, 'a' is an index indicating the distance between the azimuth distances measured by the two ranging devices and the predicted values ​​by the fusion kinematic model. Formula (11) reflects the iterative update process of the temporary weights of each sub-model. Formula (11) normalizes this process, and the sum of the weights of each sub-model should equal 1. This is the updated, unnormalized temporary weight vector. This is the normalized temporary weight.

[0105] Step 5: Temporary weights With Markov state transition matrix P k-1 Multiplying them yields the final corrected sub-model weights at time k. And normalize it;

[0106] Sub-model weights The calculation process is as shown in formula (14):

[0107]

[0108]

[0109] Where μ k P represents the final sub-model weights after Markov state transition matrix operations and normalization. k This is the Markov state transition matrix;

[0110]

[0111] Both are initialized to

[0112]

[0113] As the calculation proceeds, the weight is rapidly iterated.

[0114] Step 6: Based on the updated sub-model weights at time k With each sub-model at time k-1 Weight For the Markov state transition matrix P k-1 After correction, P is obtained. k It will be used at time k+1;

[0115] At this point, one loop is completed. Error feedback ensures the short-step prediction accuracy of the interactive multi-model system. Meanwhile, in step 3, false detections caused by millimeter-wave radar / machine vision cameras can be identified and discarded.

[0116] The Markov state transition matrix is ​​modified according to the current and historical weight coefficients of each sub-model, as shown in formulas (15) and (16).

[0117]

[0118]

[0119] in Let i be the probability of model i at the current time. Let be the probability of model i at the previous time step. Let be the element in the i-th row and j-th column of the Markov state transition matrix.

[0120] Example 2

[0121] In an autonomous vehicle equipped with millimeter-wave radar and machine vision cameras, the method described in Example 1 is used to accurately track dangerous vehicles ahead. The method is as follows:

[0122] Step 1: Use the bubble sort algorithm to select the nearest vehicles measured by the millimeter-wave radar and machine vision camera as the targets of interest to be tracked, i.e., the target objects.

[0123] Step 2: On the autonomous vehicle in operation, after the integrated inertial navigation system is installed for vehicle positioning, the real-time latitude and longitude coordinates of the vehicle can be obtained. After unifying the coordinates of the millimeter-wave radar and the machine vision camera, the coordinate system of the moving camera / radar needs to be transformed into the absolutely stationary geodetic coordinate system through a coordinate transformation method. The coordinate transformation method is shown in formulas (17)-(21).

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] where x W y W v w θ W For the positioning of this vehicle in global coordinates, These represent the coordinates, velocity, and azimuth of the target of interest in the millimeter-wave radar coordinate system, respectively, while x, y, vx, vy, and θ are the state values ​​of the target of interest as measured by the radar in the world coordinate system.

[0130] The acceleration and angular velocity of the target of interest in the global coordinate system are obtained by formula (22).

[0131]

[0132] Thus, the state vector X of the target vehicle is obtained. R =[x RRRRRR The vehicle's initial state is based on the first measurement by the millimeter-wave radar.

[0133] Step 3: Construct four kinematic models as shown in formulas (1)-(8).

[0134] Step 4: Based on the weighted fusion kinematic model of the target object at time k-1, predict its state at time k (including coordinates, heading angle, acceleration, and angular velocity in the global environment). The motion of the target object is considered as a weighted fusion of four kinematic models. This represents the predicted future state of the target object at time k based on the fused kinematic model at time k-1. Initial weights μ k Let it be μ0. μ0 = [0.4 0.1 0.4 0.1].

[0135] Step 5: At time k, the current position of the target object is determined by both millimeter-wave radar and machine vision camera. and Kinematic model fused with time k-1 The predicted values ​​are compared, and the one with the closest Euclidean distance at time k is taken as the reliable measurement. A positive value for 'a' indicates that the fused kinematics model is closer to the machine vision measurement. If 'a' exceeds a certain positive threshold, it means that the millimeter-wave radar measurement deviates significantly, indicating that the millimeter-wave radar has made a false detection, and vice versa.

[0136] Step 5. If a>0 in Step 4, then the machine vision measurement is reliable; if a<0, then the millimeter-wave radar measurement is reliable. By calculating the Euclidean distance between the predicted value of each sub-model at time k-1 and the reliable measurement value at time k, the weight of each sub-model is updated according to formulas (12) and (13). After correcting the weight of each sub-model, the weights of each sub-model are normalized so that their sum is 1 to obtain the temporary weights.

[0137] Step 6. Multiply the temporary weights by the probability transition matrix according to formula (14) to obtain the final corrected model weights at time k, and then normalize them.

[0138] Step 7. Correct the Markov state transition matrix based on the updated sub-model weights at time k and the probability weights of each sub-model at time k-1. At this point, one loop is complete. Error feedback ensures the short-step prediction accuracy of the interactive multi-model system, while also identifying and discarding false detections from millimeter-wave radar / machine vision cameras, which were detected in step 3.

[0139] Step 8. Further optimization: By statistically analyzing the variance of the millimeter-wave radar / machine vision camera measurements in the horizontal and vertical directions, data from another sensor is selected when false detections occur; otherwise, the data from both sensors are weighted and fused to smooth the output. like Figure 3 shown

[0140]

[0141]

[0142]

[0143] k x ,k y The weights assigned to the two sensors in the x and y directions are determined by the measurement noise in the x and y directions of the machine vision camera and millimeter-wave radar sensor used in this case. The fused values ​​are more accurate and have a more concentrated distribution than the individually selected measurements. Figure 6 , Figure 7 As shown.

[0144] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for fusing backend data of a ranging device based on multi-kinematic model fusion, characterized in that, Includes the following steps: Step 1: After unifying the coordinates of the two ranging devices, the motion of the target object measured by the two devices is regarded as a weighted fusion of four kinematic models. The four kinematic models are regarded as four sub-models, namely, constant velocity model, constant acceleration model, constant rotation and velocity model, and constant rotation and acceleration model. Step 2, based on the weighted fusion kinematic model of the target object at time k-1... The initial value is used to predict the state vector at time k. Step 3: At time k, measure the current state vector X of the target object using a ranging device. k The current state vectors measured by the two ranging devices are denoted as follows: and The two results were combined with the kinematic model predictions at time k-1. For comparison, the measurement results were screened, and the results were selected. and Zhongyu The nearest Euclidean distance is taken as the reliable measurement value, and the reliable measurement value is output in this step. Step 4, with As input, calculate the prediction value of each sub-model at time k-1 individually. by Reliable measurement at time k The distance is used to update the temporary weights of each sub-model. And normalize it; Step 5, set temporary weights With Markov state transition matrix P k-1 Multiplying them yields the final corrected sub-model weights at time k. And normalize it; Step 6: Based on the updated sub-model weights at time k With each sub-model at time k-1 Weight For the Markov state transition matrix P k-1 After correction, P is obtained. k It will be used at time k+1.

2. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, The two ranging devices in step 1 are millimeter-wave radar and lidar, lidar and machine vision camera, or millimeter-wave radar and machine vision camera.

3. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, The constant velocity model in step 1 is as follows: CV k =[x k y k v k θ k 0 0] T (1) The constant acceleration model is as follows: AC k =[x k and k v k θ k to k 0] T (3) The constant rotation rate and speed model is as follows: CTRV k =[x k y k v k θ k 0 ω k ] T (5) The constant rotation rate and acceleration model is as follows: CTRA k =[x k y k v k i k a k oh k ] T (7) In formula (8): Among them, CV k CA k CTRV k CTRA k These are the state vectors described by the constant velocity model, constant acceleration model, constant speed and velocity model, and constant speed and acceleration model, respectively, where x k y k v k θ k a k ω k These represent the target's x-coordinate, y-coordinate, heading angle, acceleration, and angular velocity in UTM coordinates at time k, respectively. k To calculate the step size.

4. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, State vector X k =[x k y k v k θ k a k ω k Each represents the x-coordinate, y-coordinate, velocity, heading angle, acceleration, and angular velocity of the detected target at time k in the global coordinate system. The predicted values ​​of the weighted fusion kinematics model The result can be obtained using formula (9): This represents the predicted state at time k+1 calculated by the weighted fusion of four kinematic models at time k. to These are the weights for each sub-model.

5. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, In step 4, the temporary weights of the sub-model The update process is shown in formulas (10), (11), (12), and (13): For the target x and y coordinates predicted by the fusion model, The x and y coordinates of the target measured by the ranging device. Manager, 'a' is an index indicating the distance between the measured positions of the two ranging devices and the predicted values ​​by the fusion kinematic model. Formula (11) reflects the iterative update process of the temporary weights of each sub-model. Formula (11) normalizes it, and the sum of the weights of each sub-model should be equal to 1. This is the updated, unnormalized temporary weight vector. This is the normalized temporary weight.

6. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, In step 5, the sub-model weights The calculation process is as shown in formula (14): Where μ k P represents the final sub-model weights after Markov state transition matrix operations and normalization. k This is the Markov state transition matrix; 7. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, In step 6, the Markov state transition matrix is ​​corrected according to the current and historical weight coefficients of each sub-model, as shown in formulas (15) and (16). in Let i be the probability of model i at the current time. Let be the probability of model i at the previous time step. Let be the element in the i-th row and j-th column of the Markov state transition matrix.

8. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 1, characterized in that, The process also includes step 7, where, if neither ranging device has detected a false detection, the two ranging devices are weighted and fused to smooth their output.

9. The data fusion method for the backend of a ranging device based on multi-kinematic model fusion as described in claim 8, characterized in that, The weighted fusion process is as follows: k x k y The weights are assigned to the two ranging devices in the xy direction.

10. The data fusion method for the back-end of a ranging device based on multi-kinematic model fusion as described in any one of claims 1-9, applied in intelligent driving, where a dangerous vehicle ahead is tracked as a target object, and reliable measurement values ​​are obtained. or Input to the intelligent driving control system.

Images