A high-precision map data association method based on joint probability

Abstract

The application discloses a high-precision map data association method based on joint probability, and comprises the following steps: step one, parameterization of high-precision map landmarks and vehicle detection landmarks; step two, setting an association threshold according to the current position of the vehicle; step three, calculating joint data association probability; step four, establishing a dynamic joint probability matrix; and step five, time sequence correction. The application determines a method for correctly matching observation landmarks and map landmarks by an intelligent vehicle by introducing joint probability among the semantic similarity, local space similarity and global structure similarity between landmarks, and can accurately and robustly position the intelligent vehicle by using a high-precision map, which is of great significance for improving the reliability and safety of an intelligent driving system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle positioning and information technology, and relates to a data association method for high-precision map positioning, specifically a high-precision map data association method based on joint probability. Background Technology

[0002] With the rapid development of intelligent vehicles and autonomous driving technologies, high-precision map positioning is crucial for achieving safe and efficient autonomous driving. In specific driving scenarios, when sensor-based positioning fails, the location information of landmarks in high-precision maps can provide accurate and robust vehicle location output. A key challenge in high-precision map-based positioning is accurately and in real-time establishing data associations (i.e., matching vehicle-observed landmarks with map landmarks). Incorrect data association can lead to inaccurate positioning, impacting the safety and reliability of intelligent driving systems and increasing the risk of traffic accidents.

[0003] Traditional data association methods, which introduce landmark semantic features and are based on the nearest neighbor method, have partially solved the problem of data association ambiguity. However, when the driving scenario is more complex, the nearest neighbor method is severely affected by landmark duplication and occlusion, and the problem of data association ambiguity still exists. Summary of the Invention

[0004] To address the problem that traditional nearest neighbor-based data association methods are severely affected by landmark duplication and occlusion in complex driving scenarios, leading to inaccurate positioning and consequently impacting the safety and reliability of intelligent driving systems, this invention provides a high-precision map data association method based on joint probability. This method introduces a joint probability among semantic similarity, local spatial similarity, and global structural similarity among landmarks to determine the correct matching between observed landmarks and map landmarks for intelligent vehicles. This enables accurate and robust positioning of intelligent vehicles using high-precision maps, which is of great significance for improving the reliability and safety of intelligent driving systems.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A high-precision map data association method based on joint probability includes the following steps:

[0007] Step 1: Parameterization of high-precision map landmarks and vehicle detection landmarks:

[0008] Step 11: Based on the high-precision map, set the semantic categories of the landmarks included in the map;

[0009] Steps 1 and 2: Classify the landmarks detected by the intelligent vehicles according to the semantics of the map landmarks;

[0010] Step 13: Parameterize the map landmarks and vehicle observation landmarks, assuming there are a total of One observation landmark, The map landmarks are represented by the following parametric equations:

[0011]

[0012] In the formula, The set of landmarks detected by the vehicle; A collection of map landmarks; For landmarks, the two-dimensional coordinates of the landmark in a high-precision map. The semantic category of the landmark;

[0013] Step 2: Set associated thresholds based on the vehicle's current location:

[0014] Step 2: Generate a pair of landmark points to be matched within a 50m range forward of the current vehicle's location;

[0015] Step 22: Generate an association threshold for each map point. The association threshold is an ellipse with the major axis along the vehicle's direction of travel. :

[0016]

[0017] In the formula, and For the first time during the operation of intelligent vehicles Two-dimensional coordinates of the map landmark. For the threshold;

[0018] Steps 2 and 3: If a detected landmark is not within the association threshold of map landmarks, then set its matching probability to the minimum value. If a detected landmark is within the association threshold of map landmarks, then proceed to step 3.

[0019] Step 3: Calculate the association probability of the joint data:

[0020] definition For a certain frame, the first The first detection landmark and the first The probability of establishing a match between map landmarks, given a definite data association. The set of landmarks detected by the vehicle , Assuming semantic similarity between the landmarks to be matched Global structural similarity Similarity with local space They are independent of each other, and the joint association probability is calculated based on the product of the above similarities:

[0021]

[0022] Step 4: Establish the dynamic joint probability matrix:

[0023] Step 41: Establish a dynamic joint probability matrix for each data association. , ;

[0024] Step 42: Combine the dynamic joint probability matrix Each column is normalized, and the data association of the current frame is determined according to the following maximum likelihood principle:

[0025]

[0026] In the formula, For the landmark matching relationship to be determined, i.e. the first The first detection landmark and the first Establish matching between map landmarks. Refers to the dynamic joint probability matrix The Middle All elements of the column, Refers to the dynamic joint probability matrix The Middle All elements of the row;

[0027] Step 5, Timing Correction:

[0028] Record the dynamic joint probability matrix at the current time. ,when A certain correlation probability When previous frames are also established, a smoothing correction is performed based on the association probability calculated at the previous time step. The specific correction formula is as follows:

[0029]

[0030] In the formula, It is the time-series corrected association probability. To correct the time length, As a smoothing factor, For the front The correlation probability at any given time;

[0031] At this point, the dynamic joint probability matrix at the current moment has been corrected through time-series smoothing, and then the correct data association is selected according to step four.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] This invention addresses vehicle localization using high-precision maps. Considering the inter-frame temporal relationship between vehicle-observed landmarks and map landmarks, it proposes a high-precision map data association method based on joint probability, incorporating the joint probabilities of semantic similarity, local spatial similarity, and global structural similarity. Compared to the commonly used nearest-neighbor-based data association methods, this method, considering joint probability, minimizes ambiguity and yields more accurate and robust localization results. This positively impacts the reliability and safety of intelligent driving systems, reducing the risk of traffic accidents involving autonomous vehicles. Attached Figure Description

[0034] Figure 1 This is the overall flowchart of the high-precision map data association method based on joint probability of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0036] This invention provides a high-precision map data association method based on joint probability, such as... Figure 1 As shown, the method includes the following steps:

[0037] Step 1: Parameterization of high-precision map landmarks and vehicle detection landmarks:

[0038] Based on the landmark detection results obtained from the intelligent vehicle's detection terminal and the equipped high-precision map, the semantic categories of the landmarks included in the map (such as traffic lights, traffic signs, streetlights, etc.) are first set. The landmarks detected by the intelligent vehicle should also be classified according to the map's landmark semantics. Then, the map landmarks and the vehicle-observed landmarks are parameterized. Assume there are a total of... One observation landmark, The map landmarks are represented by the following parametric equations:

[0039]

[0040] In the formula, The set of landmarks detected by the vehicle; A collection of map landmarks; For landmarks, the two-dimensional coordinates of the landmark in a high-precision map. The semantic category of the landmark.

[0041] Step 2: Set associated thresholds based on the vehicle's current location:

[0042] Generate a pair of landmark points to be matched within a 50m range forward of the current vehicle position. Simultaneously, generate an association threshold for each map point; the association threshold is an ellipse with the major axis along the vehicle's direction of travel. :

[0043]

[0044] In the formula, and For the first time during the operation of intelligent vehicles Two-dimensional coordinates of map landmarks, threshold The size of the association threshold is determined and should be set according to the semantic category of different landmarks.

[0045] If a detected landmark is not within the associated threshold of map landmarks, its matching probability can be set to a minimum. Therefore, the following method description assumes the detection of landmarks. At map landmarks Within the associated threshold.

[0046] Step 3: Calculate the association probability of the joint data:

[0047] definition For a certain frame, the first The first detection landmark and the first The probability of establishing a match between map landmarks. Given a defined data association. The set of landmarks detected by the vehicle , Assuming semantic similarity between the landmarks to be matched Global structural similarity Similarity with local space They are mutually independent. Therefore, the joint association probability can be calculated based on the product of the above similarities:

[0048]

[0049] Due to the discreteness of landmarks, semantic similarity It is modeled as a probability quality function. When the detected landmark has the same semantic type as the map landmark, the semantic similarity equals the average accuracy of the intelligent vehicle detection system. Otherwise, the semantic similarity equals the average false detection rate.

[0050] Local spatial similarity Inheriting the concept of nearest neighbors, it obtains the coordinates by calculating the difference between the coordinates of the landmarks to be matched. The calculation formula is as follows:

[0051]

[0052] In the formula, , For testing landmarks With map landmarks Two-dimensional coordinates, To detect the covariance matrix of landmark coordinates, it can be obtained by measuring the error of a high-precision map beforehand.

[0053] global structural similarity By introducing the geometric centers of the detection landmarks and map landmarks, the arrangement structure between landmarks is utilized. First, all detection landmarks are calculated. With map landmarks geometric center , Then, global similarity is obtained by calculating the coordinate difference between the coordinates to be matched and the geometric center. The formula is:

[0054]

[0055]

[0056]

[0057] In the formula, , For testing landmarks With map landmarks The global structure metric. It is a hyperparameter used to weight the structural similarity between the current landmark and other landmarks; in this method, it is set to 0.8.

[0058] By following the steps above, multiplying semantic similarity, global structural similarity, and local spatial similarity, we can calculate the association probability between all detected landmarks and map landmarks. .

[0059] Step 4: Establish the dynamic joint probability matrix:

[0060] To improve the computational and iterative efficiency of the data association method, a dynamic joint probability matrix is ​​first established for each data association. Its size depends on the number of currently detected landmarks and map landmarks, and its elements are the association probabilities calculated in step three, i.e. The dynamic joint probability matrix Each column is normalized due to the dynamic joint probability matrix. The size and elements of the array change continuously between frames and are not necessarily square matrices. We determine the data associations of the current frame based on the following maximum likelihood principle:

[0061]

[0062] In the formula, For the landmark matching relationship to be determined, i.e. the first The first detection landmark and the first Establish a match between map landmarks. Refers to the dynamic joint probability matrix The Middle All elements of the column, Refers to the dynamic joint probability matrix The Middle All elements of the row.

[0063] Step 5, Timing Correction:

[0064] To further improve the robustness of the data association algorithm, the temporal characteristics of the intelligent vehicle's driving process are utilized to refine the dynamic joint probability matrix calculated in step four. Corrections are made. The current landmark detector exhibits false detections, causing noise in the dynamic joint probability matrix, which may lead to erroneous data associations. These erroneous data associations are difficult to establish in previous moments. Therefore, all data associations established in step four can be corrected using data associations established in previous moments. Specifically, the dynamic joint probability matrix at the current moment is recorded. ,when A certain correlation probability When previous frames are also established, a smoothing correction is performed based on the association probabilities calculated at the previous time step. The specific correction formula is as follows:

[0065]

[0066] In the formula, It is the correlation probability after time-series correction. To correct the time length, that is, starting from the current moment, the previous time... Each frame calculates the detection landmarks. With map landmarks The probability of association between them. The smoothing factor is set to 0.9 in this method. For the front The correlation probability at time point.

[0067] At this point, the dynamic joint probability matrix at the current moment has been corrected through time-series smoothing, and the final correct data association can be selected according to step four.

[0068] Example 1:

[0069] Step 1: Parameterization of high-precision map landmarks and vehicle observation landmarks:

[0070] Based on the high-precision map equipped by the intelligent vehicle, the semantic categories of landmarks included in the map are first set. The semantic categories of landmarks used in this embodiment are shown in Table 1, and the parameterization results of map landmarks within 50m in front of the vehicle in frame 507 are shown in Table 2.

[0071] Table 1. Basis for semantic classification of landmarks in high-precision maps

[0072]

[0073] Table 2. Parameterization Results of High-Precision Map Landmarks

[0074]

[0075] A camera was selected as the detection device, and the parameterization results of the detected landmarks in frame 507 were obtained using a semantic segmentation algorithm based on neural networks, as shown in Table 3.

[0076] Table 3. Parameterization Results of Vehicle Observation Landmarks

[0077]

[0078] Step 2: Set associated thresholds based on the vehicle's current location:

[0079] A pair of landmark points to be matched is generated within a 50m range forward of the current vehicle position. Simultaneously, an association threshold is generated for each map point; the association threshold is an ellipse with the major axis along the vehicle's direction of travel. The ellipses generated around the three map points shown in Table 2 are as follows:

[0080]

[0081]

[0082]

[0083] In the formula, This is the standard equation of an ellipse.

[0084] Step 3: Calculate the association probability of the joint data:

[0085] The semantic similarity, global structural similarity, and local spatial similarity between map landmarks and observed landmarks are calculated. Finally, the product of these three similarities yields the association probability between all detected landmarks and map landmarks. The results are summarized in Table 4.

[0086] Table 4 Summary of Association Probabilities of Joint Data

[0087]

[0088] Step 4: Establish the dynamic joint probability matrix:

[0089] A dynamic joint probability matrix is ​​established for each data association. Its size depends on the number of currently detected landmarks and map landmarks, and its elements are the association probabilities calculated in step three, i.e. After normalization, the dynamic joint probability matrix is ​​obtained. as follows:

[0090]

[0091] Based on the dynamic joint probability matrix The resulting data association set is That is, the observation landmark 1 matches the map landmark 2; the observation landmark 2 matches the map landmark 3; and the observation landmark 3 matches the map landmark 1.

[0092] Step 5, Timing Correction:

[0093] To further improve the robustness of the data association algorithm, the temporal characteristics of the intelligent vehicle's driving process are utilized to calculate the joint probability matrix for five frames (502-506). Make corrections. The corrected probability matrix. as follows:

[0094]

[0095] It can be seen that the data association set obtained from the corrected probability matrix is ​​as follows: The initial data association error stemmed from a misidentification of a "traffic light" as an "unknown pole" by the detector in frame 507. After corrections were made in the previous five frames, observed landmark 3 and map landmark 4 were correctly matched. This demonstrates that considering joint probability in data association minimizes ambiguity and yields more accurate and robust results.

[0096] Example 2:

[0097] This embodiment tests the proposed data association method for high-precision map positioning on the KAIST autonomous driving public dataset. To quantitatively evaluate positioning performance, the mean absolute error (MAE) relative to the actual trajectory and the smoothness metric (S) are selected as indicators. The formulas for calculating MAE and S are as follows:

[0098]

[0099]

[0100] In the formula, For the number of times to locate, For the first The location result output by this method For the first The true value of the second positioning result.

[0101] Meanwhile, the method of this invention was compared with the most advanced open-source high-definition map data association and positioning method (SMMPDA). The experimental results are shown in Table 5.

[0102] Table 5 Comparison of localization results on real urban scene sequences

[0103]

[0104] As shown in Table 5, the MAE (unit: meters) and S (unit: meters) of the positioning trajectory output by the method proposed in this invention are improved by 35.93% and 38.58% respectively compared with the most advanced open source high-precision map positioning algorithm. The positioning accuracy and robustness are greatly improved, and the maximum positioning error of the algorithm is also significantly suppressed.

Claims

1. A high-precision map data association method based on joint probability, characterized in that... The method includes the following steps: Step 1: Parameterization of high-precision map landmarks and vehicle detection landmarks: Step 11: Based on the high-precision map, set the semantic categories of the landmarks included in the map; Steps 1 and 2: Classify the landmarks detected by the intelligent vehicles according to the semantics of the map landmarks; Step 13: Parameterize the map landmarks and vehicle observation landmarks, assuming there are a total of One observation landmark, The map landmarks are represented by the following parametric equations: In the formula, The set of landmarks detected by the vehicle; A collection of map landmarks; For landmarks, the two-dimensional coordinates of the landmark in a high-precision map. The semantic category of the landmark; Step 2: Set associated thresholds based on the vehicle's current location: Step 2: Generate a pair of landmark points to be matched within a 50m range forward of the current vehicle's location; Step 22: Generate an association threshold for each map point. The association threshold is an ellipse with the major axis along the vehicle's direction of travel. : In the formula, and For the first time during the operation of intelligent vehicles Two-dimensional coordinates of the map landmark. For threshold; Steps 2 and 3: If a detected landmark is not within the association threshold of map landmarks, then set its matching probability to the minimum value. If a detected landmark is within the association threshold of map landmarks, then execute step 3. Step 3: Calculate the association probability of the joint data: definition For a certain frame, the first The first detection landmark and the first The probability of establishing a match between map landmarks, given a definite data association. The set of landmarks detected by the vehicle , Assuming semantic similarity between the landmarks to be matched Global structural similarity Similarity with local space They are independent of each other, and the joint association probability is calculated based on the product of the above similarities: Global structural similarity The calculation formula is: In the formula, , These are the detection landmarks. With map landmarks The geometric center, , These are the detection landmarks. With map landmarks The global structure metric. It is a hyperparameter used to weight the structural similarity between the current landmark and other landmarks; Step 4: Establish the dynamic joint probability matrix: Step 41: Establish a dynamic joint probability matrix for each data association. , ; Step 42: Combine the dynamic joint probability matrix Each column is normalized, and the data association of the current frame is determined according to the following maximum likelihood principle: In the formula, For the landmark matching relationship to be determined, i.e. the first The first detection landmark and the first Establish matching between map landmarks. Refers to the dynamic joint probability matrix The Middle All elements of the column, Refers to the dynamic joint probability matrix The Middle All elements of the row; Step 5, Timing Correction: Record the dynamic joint probability matrix at the current time. ,when A certain correlation probability When previous frames are also established, a smoothing correction is performed based on the association probability calculated at the previous time step. The specific correction formula is as follows: In the formula, It is the time-series corrected association probability. To correct the time length, As a smoothing factor, For the front The correlation probability at any given time; At this point, the dynamic joint probability matrix at the current moment has been corrected through time-series smoothing, and then the correct data association is selected according to step four.

2. The high-precision map data association method based on joint probability according to claim 1, characterized in that... In step three, due to the discreteness of landmarks and semantic similarity... Modeled as a probability quality function, semantic similarity equals the average accuracy of the intelligent vehicle detection system when the detected landmark has the same semantic type as the map landmark; otherwise, semantic similarity equals the average false detection rate.

3. The high-precision map data association method based on joint probability according to claim 1, characterized in that... In step three, local spatial similarity It is obtained by calculating the coordinate difference between the landmark points to be matched, and the calculation formula is as follows: In the formula, , These are the detection landmarks. With map landmarks Two-dimensional coordinates, This is used to detect the covariance matrix of landmark coordinates.

Images