A vehicle trajectory map matching method integrating vehicle image semantics

Through the hidden Markov model and convolutional neural network fusion of on-board image semantic information, the problem of insufficient positioning accuracy of the existing map matching algorithm in high-precision map updates is solved, and high-precision trajectory correction and high-time efficiency updates are achieved.

CN114964272BActive Publication Date: 2024-11-01WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210491343.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-07
Publication Date
2024-11-01
Estimated Expiration
2042-05-07

AI Technical Summary

Technical Problem

The existing map matching algorithm has low positioning accuracy in high-precision map crowdsourcing mode updates, which cannot meet the high-precision and high-timed data update requirements, especially in complex road scenarios, which is difficult to accurately correct the vehicle trajectory position.

Method used

The multi-path output algorithm based on the Hidden Markov model combined with the semantic information of on-board images is used to extract the road scene, lane number and diversion line semantic information through convolutional neural network, and use the entropy weight method to fuse this information to achieve the synthesis of candidate trajectories. Evaluate to obtain the trajectory matching results for the maximum likelihood.

Benefits of technology

It improves the high-precision position correction ability of vehicle trajectory data, breaks through the accuracy bottleneck of map matching algorithm, enhances the utilization rate and value of crowdsourcing data of high-precision maps, and provides support for high-time efficiency updates of high-precision maps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114964272B_ABST
    Figure CN114964272B_ABST
Patent Text Reader

Abstract

The present invention discloses a vehicle trajectory map matching method integrating vehicle-mounted image semantics. First, a multi-path output algorithm based on a hidden Markov model is proposed, which can extract multiple candidate trajectories including the actual driving trajectory of the vehicle according to the spatiotemporal characteristics of the current trajectory data; secondly, a convolutional neural network model is constructed to extract semantic information for identifying vertical parallel roads and horizontal parallel roads from the vehicle-mounted image, and a quantification method for the vehicle-mounted image semantics is designed; finally, based on the entropy weight method, a comprehensive evaluation of the candidate trajectories is achieved through the quantified vehicle-mounted image semantics, and the maximum likelihood candidate trajectory is extracted as the map matching result of the vehicle trajectory. The present invention promotes the effective integration of vehicle-mounted images and map matching algorithms, can effectively improve the utilization rate and value of high-precision map crowdsourcing data, and provides method support for the position correction of high-precision map crowdsourcing vehicle trajectory data and the high-timeliness update of high-precision maps.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a vehicle track map matching method integrating vehicle-mounted image semantics, and belongs to the technical field of high-precision map crowdsourcing mode data updating. Background Art

[0002] High-precision maps are one of the infrastructure systems for future smart cars. They can provide high-precision positioning information, auxiliary environmental perception, path decision planning, cloud-vehicle-road collaborative interaction and other functions for autonomous driving cars. Crowdsourcing mode data update is a necessary way to achieve real-time update of high-precision maps. The vehicle images and vehicle trajectory data obtained by crowdsourcing vehicles can be used to obtain and update static road data such as road elements and ancillary facilities. At present, due to the low positioning accuracy and poor data quality of crowdsourcing trajectory data, the crowdsourcing mode update of high-precision maps has a low degree of automation and cannot meet the needs of high-precision and high-timeliness data update. Map matching algorithm is an effective method to achieve trajectory data position correction. It is often used for online vehicle navigation and offline data correction. It is also one of the core algorithms for high-precision map crowdsourcing trajectory data position correction. The map matching algorithm only integrates the spatiotemporal characteristics of trajectory data and solves the problem under the premise of assuming that the vehicle travels the shortest or best path. This makes the map matching algorithm itself an ideal solution, which is difficult to meet the requirements of trajectory position accuracy for high-precision map data collection and update. Considering that on-board images can reflect the environmental information of the vehicle and can be used to infer the complex driving behavior of the vehicle and the specific driving position of the vehicle in complex road scenes, how to effectively integrate the semantic information of on-board images into the map matching algorithm and break through the accuracy bottleneck of the map matching algorithm in crowdsourcing trajectory data correction is the main problem solved by the present invention. Summary of the invention

[0003] The purpose of the present invention is to provide a vehicle trajectory map matching method that integrates the semantics of on-board images, and to achieve high-precision position correction of vehicle trajectory data based on full use of the spatiotemporal characteristics of trajectory data collected by high-precision map crowdsourcing vehicles and the semantic characteristics of on-board images.

[0004] The present invention proposes a vehicle trajectory map matching method that integrates the semantics of vehicle-mounted images: first, a multi-path output algorithm based on a hidden Markov model is constructed, which can extract multiple candidate trajectories that may be driven by the current trajectory data, thereby improving the recall rate of the vehicle's actual driving trajectory; secondly, semantic information for identifying vertical parallel roads and horizontal parallel roads is extracted from the vehicle-mounted image, and a quantification method for the vehicle-mounted image semantics is designed; finally, based on the entropy weight method, a comprehensive evaluation of the candidate trajectories is achieved by integrating the quantified vehicle-mounted image semantic information, and the maximum likelihood candidate trajectory is extracted as the map matching result of the vehicle trajectory. The present invention promotes the effective integration of vehicle-mounted images and map matching algorithms, provides a new idea for breaking through the accuracy bottleneck of map matching algorithms, can effectively improve the utilization rate and value of high-precision map crowdsourcing data, and provides method support for the position correction of high-precision map crowdsourcing vehicle trajectory data and the high-timeliness update of high-precision maps. The technical solution of the present invention is a vehicle trajectory map matching method that integrates the semantics of vehicle-mounted images. It mainly includes the following steps:

[0005] Step 1: Based on the multi-path output algorithm of the hidden Markov model, multiple candidate trajectories that the current trajectory may travel are solved. Specifically, it can be divided into the following steps (such as Figure 1 shown):

[0006] Step 1.1, vehicle trajectory data preprocessing, first need to group vehicle trajectories according to vehicle ID, then sort vehicle trajectory data according to collection time, and then calculate straight-line speed according to the straight-line distance and time interval between adjacent original trajectory records, use the city's maximum speed limit of 120km / h as the threshold to screen outliers, and interrupt the trajectory data at the location where the outliers appear.

[0007] Step 1.2, preprocessing of urban road data, first ensure that there is no node in the road network data that is associated with only two roads. If so, merge the two roads into the same road; secondly, use the R* tree spatial index to cache the spatial information of the road sections in the road data, and use the red-black tree attribute index to cache the attribute information such as nodes, roads and road sections in the road data to facilitate the method's need for spatial and attribute retrieval of the data; in addition, a weighted directed graph based on the A* shortest path algorithm is constructed through node and road data.

[0008] Step 1.3, traverse each trajectory point in turn. If there is no next trajectory point, go to step 1.9;

[0009] Step 1.4, obtain the candidate road segments of the current trajectory point. With the current trajectory point as the center and a certain range as the radius, all road segments within the query range are used as candidate road segments of the current trajectory point through the spatial index constructed in step 1.2. First, determine the angle between the driving direction of the trajectory point and the driving direction of the candidate road segment. If the angle exceeds 60°, discard the current candidate road segment. For each candidate road segment that meets the conditions, calculate the point-line relationship function between the current trajectory point and the candidate road segment according to formula (1); calculate the candidate point of the current trajectory point on the candidate road segment according to formula (2);

[0010]

[0011] x=x1+r*(x2-x1); y=y1+r*(y2-y1) (2)

[0012] Where A and B are the starting point and end point of the road section, (x1, y1) and (x2, y2) are the coordinates of the starting point and end point respectively, P is the current track point, and (x, y) is the coordinates of the current track point;

[0013] Step 1.5, filter and group the candidate sections. If the point-line relationship function value calculated by step 1.4 is r∈[0,1], the candidate point on the current candidate section is added to the candidate set, and the road where the current candidate point is located is recorded to the candidate road ID set; if r∈[-ε,0)∪(1,1+ε] and the distance between the trajectory point and the candidate point is less than a certain threshold, the candidate point on the current candidate section is added to the candidate set, where ε is the allowable range value of the point-line relationship function; other candidate points will be discarded.

[0014] Step 1.6, select key candidate points in the candidate set to the candidate set. If all road IDs associated with a candidate point in the candidate set do not appear in the candidate road ID set in step 1.5, add the current candidate point to the candidate set, and add all road IDs associated with the current candidate point to the candidate road ID set; if all associated road IDs appear in the candidate road ID set, compare the distances between the two candidate points and the trajectory point, and only keep the candidate point with the shortest distance in the candidate set, and move the other candidate point to the candidate set; the calculation method of the associated road is as follows: all candidate points in the candidate set are endpoints of the road segment, and the road IDs of all road segments associated with the endpoints are the associated road IDs of the current candidate point;

[0015] Step 1.7, calculate the observation probability and transition probability of all current candidate points (Viterbi algorithm), which can be divided into the following steps (such as Figure 2 shown):

[0016] Step 1.7.1, determine whether there are candidate sets and alternative sets of the previous track point records. If there are no candidate sets and alternative sets of the previous track point records, jump to step 1.7.2, otherwise jump to step 1.7.3;

[0017] Step 1.7.2, calculate the observation probability of all candidate points in the current trajectory point candidate set and the alternative set, see formula (3) for details, take the observation probability as the overall probability of the candidate point, and add the candidate point with the overall probability score to the candidate result set and the alternative result set, then jump to step 1.7.8; the calculation formula of the observation probability is as follows:

[0018]

[0019] in, Represents the i-th original trajectory point p i The jth candidate point of is the spatial position probability of the vehicle trajectory point, as shown in formula (4); is the driving direction probability of the vehicle trajectory point, as shown in formula (5); is the probability of the relationship between the vehicle trajectory point and the candidate road section, as shown in formula (6);

[0020] The distance between the vehicle trajectory point and the candidate point satisfies the normal distribution, and the spatial position probability function F of the vehicle trajectory point d The definition is as follows:

[0021]

[0022] in, represents the jth candidate point of the i-th original trajectory point, represents the distance between the i-th original trajectory point and the candidate point on its j-th candidate path, μ d and σ d Represent the mean and standard deviation of the distance respectively;

[0023] The angle between the vehicle's driving direction and the candidate road section direction also satisfies the normal distribution. The probability function of the vehicle's driving direction is F θ The definition is as follows:

[0024]

[0025] in, Represents the i-th original trajectory point p i The jth candidate point of represents the angle between the driving direction of the i-th original trajectory point and the direction of the j-th candidate road segment, μ θ and σ θ Represent the mean and standard deviation of the angles respectively;

[0026] The relationship probability between the vehicle trajectory points and the candidate road segments is defined as follows:

[0027]

[0028] in, represents the jth candidate point of the i-th original trajectory point pi, Represents the point-line relationship function value between the i-th original trajectory point and the j-th candidate road segment.

[0029] Step 1.7.3, calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set, see formula (7) for details; determine whether the ratio of the shortest path length between the candidate point pairs to the straight-line distance of the trajectory point pairs exceeds a certain threshold, if so, add the candidate point to the candidate result set, otherwise add the candidate point to the temporary result set; use the road ID chain along which the candidate point passes as the unique identifier, when adding a candidate point to the result set, if there are candidate points with the same identifier, only save the one with the highest probability score in the result set.

[0030]

[0031] in Represents the overall probability score of the tth candidate point of the i-th trajectory point. According to the description of step 1.7.2, we can know that: is the observation probability, and its definition is detailed in step 1.7.2; Candidate point To the candidate point The transition probability between them is defined in formula (8).

[0032]

[0033] in, is the probability of the shortest path between the candidate vehicle points, as shown in formula (9); is the probability of the vehicle's travel speed between candidate points, as shown in formula (10);

[0034] The probability of the shortest path between the candidate points before and after the vehicle and the straight-line distance between the trajectory points is defined as follows:

[0035]

[0036] Among them, dis(p i-1 , p i ) represents the original trajectory point p i-1 With p i The straight-line distance between Represents the i-th original trajectory point p iThe tth candidate point of Representative candidate points and The shortest path distance between them, min and max are functions for finding the minimum and maximum values ​​respectively.

[0037] The probability of the vehicle's driving speed between the candidate points before and after is defined as follows:

[0038]

[0039] in, Represents the i-th original trajectory point p i The tth candidate point of Represented by candidate points To the candidate point The probability of traveling at speeds between Represents the candidate point To the candidate point The average speed is calculated by dividing the shortest path value between two candidate points by the travel time between two original trajectory points. u represents the maximum speed limit of the current road u, and k represents the candidate point and The number of road segments that the shortest path passes through.

[0040] Step 1.7.4, calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set. The calculation formula is detailed in step 1.7.3; determine whether the ratio of the shortest path length between the candidate point pairs to the straight-line distance of the trajectory point pairs exceeds a certain threshold. If so, add the candidate point to the candidate result set, otherwise add the candidate point to the temporary result set; use the road ID chain along which the candidate point passes as the unique identifier. When adding candidate points to the result set, if candidate points with the same identifier appear, only save the item with the closest matching distance in the result set.

[0041] Step 1.7.5, obtain the associated roads of all candidate points in the current trajectory point candidate set and the candidate result set, marked as AR1, AR2, and determine whether all the associated roads of the candidate set appear in the candidate result set, that is, If yes, jump directly to step 1.7.6; otherwise, count the IDs of the associated roads that did not appear, identify them with AR, where AR∈AR2-AR1, and calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set. The calculation formula is detailed in step 1.7.3. Select the candidate points whose associated roads intersect with the above-mentioned associated roads that did not appear from the calculation results, and add the candidate points to the candidate result set; if candidate points with the same identifier appear, only save the one with the highest probability score in the candidate result set.

[0042] Step 1.7.6, again obtain the associated roads of all candidate points in the current candidate result set, marked as AR2, and at the same time obtain the associated roads of all candidate points in the previous trajectory point candidate set, marked as AR3, and determine whether all the associated roads of the previous trajectory point candidate set appear in the candidate result set, that is, If yes, jump directly to step 1.7.7; otherwise, count the IDs of the associated roads that do not appear, identify them with AR, where AR∈AR2-AR3, and select candidate points from the alternative result set where the associated roads intersect with the above-mentioned associated roads that do not appear, and move the candidate points from the alternative result set to the candidate result set; if candidate points with the same identifier appear, only save the item with the closest map matching distance in the candidate result set.

[0043] Step 1.7.7, if the candidate result set is not empty, jump directly to step 1.7.8; otherwise, add all the data in the temporary result set to the candidate result set;

[0044] Step 1.7.8, returning the candidate result of the current trajectory point and the candidate result set as the result;

[0045] Step 1.8, if the candidate result set and alternative result set returned by the Viterbi algorithm are empty, go to step 1.9, otherwise record the candidate result set and alternative result set of the current trajectory point, use them as the pre-order candidate set and pre-order alternative set of the next trajectory point, and go to step 1.3;

[0046] Step 1.9, recursive solution and candidate point screening. For each candidate point retained by the current trajectory point, recursively search its predecessor candidate point in turn, and flip the obtained candidate point chain to obtain the candidate trajectory of the current candidate point. For each candidate trajectory, use the ID chain of the path between the second candidate point and the second to last candidate point as the unique identifier, filter all candidate points, and if there are two candidate points with the same identifier, only retain the one with the highest probability score.

[0047] Step 1.10, track splicing. If the result set is empty, all candidate tracks currently being solved are added to the result set; otherwise, the candidate tracks currently being solved are spliced ​​with the tracks saved in the result set: based on the candidate tracks currently being solved, find a candidate track in the result set whose last candidate point is on the same road as the first candidate point of the current candidate track, and splice it with the current candidate track; if the above situation does not exist, select the candidate track in the result set whose distance between the last candidate point and the first candidate point of the current candidate track is the shortest, and splice it with the current candidate track; the result after splicing is retained as the multi-path output result of the current track point solution;

[0048] Step 1.11, if there are still trajectory points that have not been traversed, clear the currently recorded candidate result set and alternative result set, and go to step 1.3; otherwise, the multi-path output algorithm is solved, and go to step 2.

[0049] Step 2, vehicle image semantic extraction and quantification, can be divided into the following steps (step 1-3 detailed flow chart, as shown in Figure 3 shown):

[0050] Step 2.1, road scene semantic extraction and quantification, using convolutional neural network to classify the scene mode of the vehicle image into four categories: tunnel, viaduct without occlusion, viaduct occlusion and no occlusion; the quantification method of road scene semantics is as follows:

[0051]

[0052] Among them, α∈[0,1] is the confidence of the semantic recognition of the vehicle-mounted image, and the situations where the road scene is consistent with the candidate road attributes include: the situation where both the image and the road scene are tunnels; the situation where the image is blocked by an elevated bridge and the road scene is non-elevated and tunnel; the situation where the image is an elevated bridge without occlusion or without occlusion. If there is no elevated bridge in the road scene of all candidate points of the current trajectory point, it is considered that the attributes of all non-tunnel roads are consistent with the road scene of the current image. If there is an elevated bridge, the unobstructed scene of the elevated bridge is consistent with the attributes of all non-tunnel roads, and the unobstructed scene is consistent with the attributes of the elevated road. The score of each candidate trajectory in terms of road scene semantics is equal to the sum of the scores of all candidate points on the candidate trajectory in terms of road scene semantics;

[0053] Step 2.2, lane number semantic extraction and quantification, using convolutional neural network, identify the number of lanes in the vehicle image (ignoring the lanes on the left side of the solid yellow line and the non-motorized vehicle lane on the right), and set the roads without lane markings as 1 lane by default, and manually correct the vehicle image recognition results and urban road attributes of more than 4 lanes to 4 lanes. The quantification method of the lane number semantics is as follows:

[0054]

[0055] Among them, α∈[0,1] is the confidence of semantic recognition of vehicle-mounted images, y p ,y r are the number of lanes extracted from the corrected vehicle image and the number of lanes recorded in the urban road attributes, respectively. max and abs are functions for calculating the maximum value and absolute value, respectively. The score of each candidate trajectory in terms of the semantics of the number of lanes is equal to the sum of the scores of all candidate points on the candidate trajectory in terms of the semantics of the number of lanes.

[0056] Step 2.3, guide line semantic extraction and quantification, using convolutional neural network to identify the guide lines appearing in the vehicle image, including: no guide line, guide line appears on the left side of the vehicle, guide line appears on the right side of the vehicle, guide line appears on both sides of the vehicle. The guide line semantic quantification method is as follows:

[0057]

[0058] Among them, α∈[0,1] is the confidence of semantic recognition of vehicle-mounted images. If there is no guide line in the current vehicle-mounted image, the score of all candidate points at the current trajectory point is α; if a guide line appears in the vehicle-mounted image, each candidate point of the current trajectory point is taken as the center, and 5 consecutive candidate points are selected on each candidate trajectory to form a trajectory chain. The overlapping trajectory chains are grouped using a union-find data structure; after grouping, if there is only one trajectory chain in the group, the score of the candidate trajectory at the current trajectory point is 0.8α; for scenes where there are multiple trajectory chains in the group, the trajectory chains are sorted according to the position where the guide line appears. For example: if the current guide line appears on the left side of the vehicle, it proves that the vehicle is driving on the right side. The score of the trajectory on the right in the same group is α, and the scores of other trajectories decrease by 0.2α from right to left; if the guide line appears on both sides of the vehicle, the score of one or two candidate trajectories in the middle is α, and the scores of other trajectories decrease by 0.2α from the center to both sides. In formula (13), c j represents the number of trajectory chains in the jth group, and i represents the order of the trajectory chains in the group.

[0059] Step 3, entropy weight method evaluation and map matching result solution, can be divided into the following steps:

[0060] Step 3.1, data normalization:

[0061]

[0062] where x i,j represents the semantic score of the jth vehicle image of the i-th candidate track, min i=1 ...nx i,j and max i=1 ...nx i,j They represent the maximum and minimum values ​​of a certain type of vehicle image semantics in all candidate trajectories.

[0063] Step 3.2, weight calculation, calculate the weight of each type of vehicle image semantics in each candidate track in the corresponding indicators of all candidate tracks:

[0064]

[0065] Step 3.3, entropy calculation, calculate the entropy value of the semantic score of each type of vehicle image:

[0066]

[0067] Step 3.4, information entropy redundancy calculation, calculate the information entropy redundancy of each type of vehicle image semantics:

[0068] d j =1-e j (17)

[0069] Step 3.5, weight calculation, calculate the weight of each vehicle image semantics:

[0070]

[0071] Step 3.6, calculate the score of each candidate trajectory, and select the candidate trajectory with the highest overall score as the result of map matching:

[0072]

[0073] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: First, the present invention proposes a multi-path output algorithm based on a hidden Markov model, which can extract multiple candidate trajectories including the real driving trajectory of the vehicle according to the spatiotemporal characteristics of the current trajectory data; secondly, a convolutional neural network model is constructed to extract semantic information for identifying vertical parallel roads and horizontal parallel roads from the vehicle-mounted image, and a quantification method for the semantics of the vehicle-mounted image is designed; finally, based on the entropy weight method, a comprehensive evaluation of the candidate trajectories is realized through the quantified vehicle-mounted image semantics, and the maximum likelihood candidate trajectory is extracted as the map matching result of the vehicle trajectory. The present invention realizes the high-precision position correction of the vehicle trajectory data on the basis of making full use of the spatiotemporal characteristics of the trajectory data collected by the high-precision map crowdsourcing vehicle and the semantic characteristics of the vehicle-mounted image, promotes the effective integration of the vehicle-mounted image and the map matching algorithm, provides a new idea for breaking through the accuracy bottleneck of the map matching algorithm, can effectively improve the utilization rate and value of the high-precision map crowdsourcing data, and provides method support for the position correction of the high-precision map crowdsourcing vehicle trajectory data and the high-precision map with high timeliness update. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 : Multi-path output algorithm logic flow chart.

[0075] Figure 2 : Viterbi algorithm logic flow chart.

[0076] Figure 3 : Detailed flow chart of the vehicle trajectory map matching method integrating vehicle-mounted image semantics. DETAILED DESCRIPTION

[0077] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0078] Taking the high-precision map crowdsourcing data of Wuhan City as an example, the specific implementation process includes the following steps:

[0079] Step 1: Vehicle trajectory data preprocessing: First, the vehicle trajectories need to be grouped according to the vehicle ID, and then the vehicle trajectory data is sorted according to the collection time. The straight-line speed is calculated based on the straight-line distance and time interval between adjacent original trajectory records. The maximum speed limit of 120km / h is used as the threshold to screen outliers, and the trajectory data is interrupted at the location where the outliers appear.

[0080] Step 2: Preprocessing of navigation network data: Use red-black trees to construct attribute indexes for roads, road sections, nodes, and other information, use R* trees to construct spatial indexes for road section information, and then use node and road data to construct a weighted directed graph for the A* shortest path algorithm.

[0081] Step 3, track point traversal: if there is a next track point, execute step 4; if there is no next track point, execute step 9;

[0082] Step 4, candidate road segment acquisition: traverse each trajectory point in turn, take the current trajectory point as the center and a certain range as the radius, query all road segments within the range as candidate road segments for the current trajectory point, first determine the angle between the driving direction of the trajectory point and the driving direction of the candidate road segment, if the angle exceeds 60°, discard the current candidate road segment; for each candidate road segment that meets the conditions, calculate the point-line relationship function between the current trajectory point and the candidate road segment according to formula (1), and then solve the candidate point of the current trajectory point on the candidate road segment according to formula (2).

[0083]

[0084] x=x1+r*(x2-x1); y=y1+r*(y2-y1) (2)

[0085] Where A and B are the starting point and end point of the road section, (x1, y1) and (x2, y2) are the coordinates of the starting point and end point respectively, P is the current track point, and (x, y) is the coordinates of the current track point;

[0086] Step 5, the candidate sections are screened and grouped: if the calculated point-line relationship function value r∈[0, 1], the candidate point on the current candidate section is added to the candidate set, and the road where the current candidate point is located is recorded to the candidate road ID set; if r∈[-ε, 0)∪(1, 1+ε] and the distance between the trajectory point and the candidate point is less than a certain threshold, the candidate point on the current candidate section is added to the candidate set; other candidate points will be discarded.

[0087] Step 6, candidate point screening in the candidate set: if all road IDs associated with a candidate point in the candidate set do not appear in the road ID set associated with all candidate points in the candidate set, then add the current candidate point to the candidate set, and add all road IDs associated with the current candidate point to the candidate road ID set; if all associated road IDs appear in the candidate road ID set, then compare the distance between the candidate point and the trajectory point, and only keep the candidate point with the shortest distance in the candidate set, and move the other candidate point to the candidate set;

[0088] Step 7, Viterbi algorithm: It can be divided into the following steps:

[0089] Step 7.1, determine whether there are candidate sets and alternative sets of the previous trajectory point records, if not, proceed to step 7.2, if yes, proceed to step 7.3;

[0090] Step 7.2, calculate the observation probability of all candidate points in the current trajectory point candidate set and the alternative set according to formula (3), and use the observation probability as the overall probability of the candidate point, and add the candidate point with the overall probability score to the candidate result set and the alternative result set; then execute step 7.8;

[0091]

[0092] in, Represents the i-th original trajectory point p i The jth candidate point of is the spatial position probability of the vehicle trajectory point, as shown in formula (4); is the driving direction probability of the vehicle trajectory point, as shown in formula (5); is the probability of the relationship between the vehicle trajectory point and the candidate road section, as shown in formula (6);

[0093] The distance between the vehicle trajectory point and the candidate point satisfies the normal distribution, and the spatial position probability function F of the vehicle trajectory point d The definition is as follows:

[0094]

[0095] in, represents the jth candidate point of the i-th original trajectory point, represents the distance between the i-th original trajectory point and the candidate point on its j-th candidate path, μ d and σ d Represent the mean and standard deviation of the distance respectively;

[0096] The angle between the vehicle's driving direction and the candidate road section direction also satisfies the normal distribution. The probability function of the vehicle's driving direction is Fθ The definition is as follows:

[0097]

[0098] in, Represents the i-th original trajectory point p i The jth candidate point of represents the angle between the driving direction of the i-th original trajectory point and the direction of the j-th candidate road segment, μ θ and σ θ Represent the mean and standard deviation of the angles respectively;

[0099] The relationship probability between the vehicle trajectory points and the candidate road segments is defined as follows:

[0100]

[0101] in, Represents the i-th original trajectory point p i The jth candidate point of Represents the point-line relationship function value between the i-th original trajectory point and the j-th candidate road segment.

[0102] Step 7.3, calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set according to formula (7), and determine whether the ratio of the shortest path length between the candidate point pairs to the straight-line distance of the trajectory point pairs exceeds a certain threshold. If yes, add the candidate point to the candidate result set, otherwise add the candidate point to the temporary result set; use the road ID chain along which the candidate point passes as the unique identifier. When adding candidate points to the result set, if there are candidate points with the same identifier, only save the one with the highest probability score in the result set;

[0103]

[0104] in Represents the overall probability score of the tth candidate point of the i-th trajectory point. According to the description of step 7.2, we can know that: is the observation probability, which is defined in step 7.2; Candidate point To the candidate point The transition probability between them is defined in formula (8).

[0105]

[0106] in, is the probability of the shortest path between the candidate vehicle points, as shown in formula (9); is the probability of the vehicle's travel speed between candidate points, as shown in formula (10);

[0107] The probability of the shortest path between the candidate points before and after the vehicle and the straight-line distance between the trajectory points is defined as follows:

[0108]

[0109] Among them, dis(p i-1 , p i ) represents the original trajectory point p i-1 With p i The straight-line distance between Represents the i-th original trajectory point p i The tth candidate point of Representative candidate points and The shortest path distance between them, min and max are functions for finding the minimum and maximum values ​​respectively.

[0110] The probability of the vehicle's driving speed between the candidate points before and after is defined as follows:

[0111]

[0112] in, Represents the i-th original trajectory point p i The tth candidate point of Represented by candidate points To the candidate point The probability of traveling at speeds between Represents the candidate point To the candidate point The average speed is calculated by dividing the shortest path value between two candidate points by the travel time between two original trajectory points. u represents the maximum speed limit of the current road u, and k represents the candidate point and The number of road segments that the shortest path passes through.

[0113] Step 7.4, calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set, and determine whether the ratio of the shortest path length between the candidate point pairs to the straight-line distance of the trajectory point pairs exceeds a certain threshold. If so, add the candidate point to the candidate result set, otherwise add the candidate point to the temporary result set; use the road ID chain along which the candidate point passes as the unique identifier. When adding a candidate point to the result set, if a candidate point with the same identifier appears, only save the item with the closest matching distance in the result set.

[0114] Step 7.5, obtain the associated roads of all candidate points in the current trajectory point candidate set and the candidate result set, marked as AR1, AR2, and determine whether all the associated roads of the candidate set appear in the candidate result set, that is, If yes, directly execute step 7.6; otherwise, count the IDs of the associated roads that did not appear, identify them with AR, where AR∈AR2-AR1, and calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set, select the candidate points whose associated roads intersect with the above-mentioned associated roads that did not appear from the calculation results, and add the candidate points to the candidate result set; if candidate points with the same identifier appear, only save the one with the highest probability score in the candidate result set.

[0115] Step 7.6, again obtain the associated roads of all candidate points in the current candidate result set, marked as AR2, and at the same time obtain the associated roads of all candidate points in the previous trajectory point candidate set, marked as AR3, and determine whether all the associated roads of the previous trajectory point candidate set appear in the candidate result set, that is, If yes, directly execute step 7.7; otherwise, count the IDs of the associated roads that do not appear, identify them with AR, where AR∈AR2-AR3, and select candidate points from the alternative result set where the associated roads intersect with the above-mentioned associated roads that do not appear, and move the candidate points from the alternative result set to the candidate result set; if candidate points with the same identifier appear, only save the item with the closest map matching distance in the candidate result set.

[0116] Step 7.7, if the candidate result set is not empty, directly execute step 7.8; otherwise, add all the data in the temporary result set to the candidate result set;

[0117] Step 7.8, returning the candidate result of the current trajectory point and the candidate result set as the result;

[0118] Step 8, Viterbi result recording: If the candidate result set and alternative result set returned by the Viterbi algorithm are empty, execute step 9; otherwise, record the candidate result set and alternative result set of the current trajectory point and use them as the pre-order candidate set and pre-order alternative set of the next trajectory point, and execute step 3;

[0119] Step 9, recursive solution and candidate point screening: for each candidate point retained by the current trajectory point, recursively solve its previous candidate point in turn, and flip the obtained candidate point chain to obtain the candidate trajectory of the current candidate point; for each candidate trajectory, use the ID chain of the path between the second candidate point and the second to last candidate point as the unique identifier, filter all candidate points, and if there are two candidate points with the same identifier, only retain the one with the highest probability score;

[0120] Step 10, track splicing: if the result set is empty, all candidate tracks currently being solved are added to the result set; otherwise, the candidate tracks currently being solved are spliced ​​with the tracks saved in the result set: based on the candidate tracks currently being solved, find a candidate track in the result set whose last candidate point is on the same road as the first candidate point of the current candidate track, and splice it with the current candidate track; if the above situation does not exist, select the candidate track in the result set whose distance between the last candidate point and the first candidate point of the current candidate track is the shortest, and splice it with the current candidate track; the result after splicing is retained as the multi-path output result of the current track point solution;

[0121] Step 11, multi-path output algorithm judgment: if there are still trajectory points that have not been traversed, clear the currently recorded candidate result set and alternative result set, and go to step 3; otherwise, the multi-path output algorithm solution is completed, and go to step 12.

[0122] Step 12, road scene semantic extraction and quantification: the acquired Baidu panorama data is annotated with road scene semantic information, and a convolutional neural network is trained based on the annotated panorama data; then the trained convolutional neural network model is used to extract the road scene semantics of each vehicle-mounted image, and finally the road scene semantics is quantified, and the sum of the road scene semantic scores of all candidate points on the candidate trajectory is taken as the total road scene semantic score of the current candidate trajectory; the vehicle-mounted image scene modes are divided into four categories: tunnel, viaduct unobstructed, viaduct obstructed, and unobstructed; the road scene semantics quantification method is as follows:

[0123]

[0124] Among them, α∈[0,1] is the confidence of the semantic recognition of the vehicle-mounted image, and the situations where the road scene is consistent with the candidate road attributes include: the situation where both the image and the road scene are tunnels; the situation where the image is blocked by an elevated bridge and the road scene is non-elevated and tunnel; the situation where the image is an elevated bridge without occlusion or without occlusion. If there is no elevated bridge in the road scene of all candidate points of the current trajectory point, it is considered that the attributes of all non-tunnel roads are consistent with the road scene of the current image. If there is an elevated bridge, the unobstructed scene of the elevated bridge is consistent with the attributes of all non-tunnel roads, and the unobstructed scene is consistent with the attributes of the elevated road. The score of each candidate trajectory in terms of road scene semantics is equal to the sum of the scores of all candidate points on the candidate trajectory in terms of road scene semantics;

[0125] Step 13, lane number semantic extraction and quantification: the acquired Baidu panorama data is annotated with lane number semantic information, and a convolutional neural network is trained based on the annotated panorama data; then the trained convolutional neural network model is used to extract the number of lanes in each vehicle image, and finally the lane number semantics is quantified, and the sum of the lane number semantic scores of all candidate points on the candidate trajectory is taken as the total lane number semantic score of the current candidate trajectory; when quantifying, roads without lane line markings are defaulted to 1 lane, and vehicle image recognition results and urban road attributes with more than 4 lanes are manually corrected to 4 lanes. The quantification method of lane number semantics is as follows:

[0126]

[0127] Among them, α∈[0,1] is the confidence of semantic recognition of vehicle-mounted images, y p ,y r are the number of lanes extracted from the corrected vehicle image and the number of lanes recorded in the urban road attributes, respectively. max and abs are functions for calculating the maximum value and absolute value, respectively. The score of each candidate trajectory in terms of the semantics of the number of lanes is equal to the sum of the scores of all candidate points on the candidate trajectory in terms of the semantics of the number of lanes.

[0128] Step 14, guide line semantic extraction and quantification: the guide line semantic information of the acquired Baidu panorama data is annotated, and the convolutional neural network is trained based on the annotated panorama data (ignoring the lane on the left side of the yellow solid line and the non-motorized lane on the right); then the trained convolutional neural network model is used to extract the guide line position of each vehicle-mounted image, and finally the guide line semantics is quantified, and the sum of the guide line semantic scores of all candidate points on the candidate trajectory is taken as the total guide line semantic score of the current candidate trajectory; the guide line recognition content includes: no guide line, guide line appears on the left side of the vehicle, guide line appears on the right side of the vehicle, and guide line appears on both sides of the vehicle. The guide line semantics quantification method is as follows:

[0129]

[0130] Among them, α∈[0,1] is the confidence of semantic recognition of vehicle-mounted images. If there is no guide line in the current vehicle-mounted image, the score of all candidate points at the current trajectory point is α; if a guide line appears in the vehicle-mounted image, each candidate point of the current trajectory point is taken as the center, and 5 consecutive candidate points are selected on each candidate trajectory to form a trajectory chain. The overlapping trajectory chains are grouped using a union-find data structure; after grouping, if there is only one trajectory chain in the group, the score of the candidate trajectory at the current trajectory point is 0.8α; for scenes where there are multiple trajectory chains in the group, the trajectory chains are sorted according to the position where the guide line appears. For example: if the current guide line appears on the left side of the vehicle, it proves that the vehicle is driving on the right side. The score of the trajectory on the right in the same group is α, and the scores of other trajectories decrease by 0.2α from right to left; if the guide line appears on both sides of the vehicle, the score of one or two candidate trajectories in the middle is α, and the scores of other trajectories decrease by 0.2α from the center to both sides. In formula (13), c j represents the number of trajectory chains in the jth group, and i represents the order of the trajectory chains in the group.

[0131] Step 15, entropy weight method evaluation and map matching result solution: normalize the scores of various types of vehicle image semantics; use the entropy weight method to calculate the weight of each type of vehicle image semantics; then calculate the weighted sum of the normalized scores of all vehicle image semantics as the final score of the candidate trajectory, and obtain the candidate trajectory with the highest score as the result of map matching.

[0132] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A vehicle trajectory map matching method integrating vehicle-mounted image semantics, characterized in that: The steps include: Step 1: Based on the multi-path output algorithm of the hidden Markov model, multiple candidate trajectories that the current trajectory may travel are solved; The specific implementation of step 1 is as follows: Step 1.1, vehicle trajectory data preprocessing, first need to group vehicle trajectories according to vehicle ID, then sort vehicle trajectory data according to collection time, and then calculate straight-line speed according to the straight-line distance and time interval between adjacent original trajectory records, use the city's maximum speed limit Nkm / h as the threshold to screen outliers, and interrupt trajectory data at the location where the outliers appear; Step 1.2, preprocessing of urban road data, first ensure that there is no node in the road network data that is associated with only two roads, if so, merge the two roads into the same road; secondly, use R* tree spatial index to cache the spatial information of road sections in the road data, and use red-black tree attribute index to cache the node, road and road section attribute information in the road data to facilitate the method's need for spatial and attribute retrieval of data; in addition, a weighted directed graph based on the A* shortest path algorithm is constructed through node and road data; Step 1.3, traverse each trajectory point in turn. If there is no next trajectory point, go to step 1.9; Step 1.4, obtain the candidate road segments of the current trajectory point. With the current trajectory point as the center and a certain range as the radius, all road segments within the query range are used as candidate road segments of the current trajectory point through the spatial index constructed in step 1.

2. First, determine the angle between the driving direction of the trajectory point and the driving direction of the candidate road segment. If the angle exceeds M°, discard the current candidate road segment. For each candidate road segment that meets the conditions, calculate the point-line relationship function between the current trajectory point and the candidate road segment according to formula (1); Calculate the candidate point of the current trajectory point on the candidate road section according to formula (2); x=x1+r*(x2-x1); y=y1+r*(y2-y1) (2) Where A and B are the starting point and end point of the road section, (x1, y1) and (x2, y2) are the coordinates of the starting point and end point respectively, P is the current track point, and (x, y) is the coordinates of the current track point; Step 1.5, filter and group the candidate sections. If the point-line relationship function value calculated by step 1.4 is r∈[0, 1], the candidate point on the current candidate section is added to the candidate set, and the road where the current candidate point is located is recorded to the candidate road ID set; if r∈[-ε, 0)∪(1, 1+ε] and the distance between the trajectory point and the candidate point is less than a certain threshold, the candidate point on the current candidate section is added to the candidate set, where ε is the allowable range value of the point-line relationship function; other candidate points will be discarded; Step 1.6, select key candidate points in the candidate set to the candidate set. If all road IDs associated with a candidate point in the candidate set do not appear in the candidate road ID set in step 1.5, add the current candidate point to the candidate set, and add all road IDs associated with the current candidate point to the candidate road ID set; if all associated road IDs appear in the candidate road ID set, compare the distances between the two candidate points and the trajectory point, and only keep the candidate point with the shortest distance in the candidate set, and move the other candidate point to the candidate set; the calculation method of the associated road is as follows: all candidate points in the candidate set are endpoints of the road segment, and the road IDs of all road segments associated with the endpoints are the associated road IDs of the current candidate point; Step 1.7, use the Viterbi algorithm to calculate the observation probability and transition probability of all current candidate points; Step 1.8, if the candidate result set and alternative result set returned by the Viterbi algorithm are empty, go to step 1.9, otherwise record the candidate result set and alternative result set of the current trajectory point, use them as the pre-order candidate set and pre-order alternative set of the next trajectory point, and go to step 1.3; Step 1.9, recursive solution and candidate point screening. For each candidate point retained by the current trajectory point, recursively search its predecessor candidate point in turn, and flip the obtained candidate point chain to obtain the candidate trajectory of the current candidate point. For each candidate trajectory, use the ID chain of the path between the second candidate point and the second to last candidate point as the unique identifier, filter all candidate points, and if there are two candidate points with the same identifier, only retain the one with the highest probability score. Step 1.10, track splicing. If the result set is empty, all candidate tracks currently being solved are added to the result set; otherwise, the candidate tracks currently being solved are spliced ​​with the tracks saved in the result set: based on the candidate tracks currently being solved, find a candidate track in the result set whose last candidate point is on the same road as the first candidate point of the current candidate track, and splice it with the current candidate track; if the above situation does not exist, select the candidate track in the result set whose distance between the last candidate point and the first candidate point of the current candidate track is the shortest, and splice it with the current candidate track; The result after splicing is retained as the multi-path output result of the current trajectory point solution; Step 1.11, if there are still trajectory points that have not been traversed, clear the candidate result set and the alternative result set of the current record, and go to step 1.3; otherwise, the multi-path output algorithm is solved, and go to step 2; Step 2: semantic extraction and quantification of vehicle images, specifically including using convolutional neural networks to classify the scene patterns of vehicle images, identify the number of lanes in the vehicle images, and identify the guide lines that appear in the vehicle images; Step 3: Based on the entropy weight method, the candidate trajectories are comprehensively evaluated through the quantized vehicle image semantics, and the candidate trajectory with the maximum likelihood is extracted as the map matching result of the vehicle trajectory.

2. A vehicle trajectory map matching method integrating vehicle-mounted image semantics as claimed in claim 1, characterized in that: The specific implementation of step 1.7 is as follows; Step 1.7.1, determine whether there are candidate sets and alternative sets of the previous track point records. If there are no candidate sets and alternative sets of the previous track point records, jump to step 1.7.2, otherwise jump to step 1.7.3; Step 1.7.2, calculate the observation probability of all candidate points in the current trajectory point candidate set and the alternative set, see formula (3) for details, take the observation probability as the overall probability of the candidate point, and add the candidate point with the overall probability score to the candidate result set and the alternative result set, then jump to step 1.7.8; The observation probability calculation formula is as follows: in, Represents the i-th original trajectory point p i The jth candidate point of is the spatial position probability of the vehicle trajectory point, as shown in formula (4); is the driving direction probability of the vehicle trajectory point, as shown in formula (5); is the probability of the relationship between the vehicle trajectory point and the candidate road section, as shown in formula (6); The distance between the vehicle trajectory point and the candidate point satisfies the normal distribution, and the spatial position probability function F of the vehicle trajectory point d The definition is as follows: in, represents the jth candidate point of the i-th original trajectory point, represents the distance between the i-th original trajectory point and the candidate point on its j-th candidate path, μ d and σ d Represent the mean and standard deviation of the distance respectively; The angle between the vehicle's driving direction and the candidate road section direction also satisfies the normal distribution. The probability function of the vehicle's driving direction is F θ The definition is as follows: in, Represents the i-th original trajectory point p i The jth candidate point of represents the angle between the driving direction of the i-th original trajectory point and the direction of the j-th candidate road segment, μ θ and σ θ Represent the mean and standard deviation of the angles respectively; The relationship probability between the vehicle trajectory points and the candidate road segments is defined as follows: in, Represents the i-th original trajectory point p i The jth candidate point of Represents the point-line relationship function value between the i-th original trajectory point and the j-th candidate road segment; Step 1.7.3, calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set, see formula (7) for details; determine whether the ratio of the shortest path length between the candidate point pairs to the straight-line distance of the trajectory point pairs exceeds a certain threshold, if yes, add the candidate point to the candidate result set, otherwise add the candidate point to the temporary result set; use the road ID chain along which the candidate point passes as the unique identifier, when adding candidate points to the result set, if there are candidate points with the same identifier, only save the one with the highest probability score in the result set; in Represents the overall probability score of the tth candidate point of the i-th trajectory point. According to the description of step 1.7.2, we can know that: is the observation probability, and its definition is detailed in step 1.7.2; Candidate point To the candidate point The transition probability between is defined in formula (8); in, is the probability of the shortest path between the candidate vehicle points, as shown in formula (9); is the probability of the vehicle's travel speed between candidate points, as shown in formula (10); The probability of the shortest path between the candidate points before and after the vehicle and the straight-line distance between the trajectory points is defined as follows: Among them, dis(p i-1 , p i ) represents the original trajectory point p i-1 With p i The straight-line distance between Represents the i-th original trajectory point p i The tth candidate point of Representative candidate points and The shortest path distance between them, min and max are functions for finding the minimum and maximum values ​​respectively; The probability of the vehicle's driving speed between the candidate points before and after is defined as follows: in, represents the tth candidate point of the ith original trajectory point pi, Represented by candidate points To the candidate point The probability of traveling at speeds between Represents the candidate point To the candidate point The average speed is calculated by dividing the shortest path value between two candidate points by the travel time between two original trajectory points. u represents the maximum speed limit of the current road u, and k represents the candidate point and The number of road segments that the shortest path passes through; Step 1.7.4, calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set. The calculation formula is detailed in step 1.7.3; determine whether the ratio of the shortest path length between the candidate point pairs to the straight-line distance between the trajectory point pairs exceeds a certain threshold. If so, add the candidate point to the candidate result set; otherwise, add the candidate point to the temporary result set; use the road ID chain along which the candidate point passes as the unique identifier. When adding candidate points to the result set, if candidate points with the same identifier appear, only save the item with the closest matching distance in the result set; Step 1.7.5, obtain the associated roads of all candidate points in the current trajectory point candidate set and the candidate result set, and determine whether all the associated roads in the candidate set appear in the candidate result set. If yes, jump directly to step 1.7.6; otherwise, count the IDs of the associated roads that do not appear, and calculate the overall probability of all candidate point pairs between the current trajectory point candidate set and the previous trajectory point candidate set. The calculation formula is detailed in step 1.7.

3. Select the candidate points whose associated roads intersect with the above-mentioned associated roads that do not appear from the calculation results, and add the candidate points to the candidate result set; if candidate points with the same ID appear, only save the item with the highest probability score in the candidate result set; Step 1.7.6, again obtain the associated roads of all candidate points in the current candidate result set, and at the same time obtain the associated roads of all candidate points in the previous trajectory point candidate set, and determine whether all the associated roads in the previous trajectory point candidate set appear in the candidate result set. If yes, directly jump to step 1.7.7; otherwise, count the IDs of the associated roads that do not appear, and select candidate points whose associated roads intersect with the above-mentioned associated roads that do not appear from the alternative result set, and move the candidate points from the alternative result set to the candidate result set; if candidate points with the same ID appear, only save the item with the closest map matching distance in the candidate result set; Step 1.7.7, if the candidate result set is not empty, jump directly to step 1.7.8; otherwise, add all the data in the temporary result set to the candidate result set; Step 1.7.8, return the candidate result of the current trajectory point and the alternative result set as the result.

3. The vehicle trajectory map matching method integrating vehicle-mounted image semantics according to claim 1, characterized in that: In step 2, a convolutional neural network is used to classify the scene modes of the vehicle images into four categories: tunnel, viaduct without occlusion, viaduct occlusion, and no occlusion. The quantification method of road scene semantics is as follows: Among them, α∈[0,1] is the confidence of the semantic recognition of the vehicle-borne image, and the situations where the road scene is consistent with the candidate road attributes include: the situation where both the image and the road scene are tunnels; the situation where the image is blocked by an elevated bridge and the road scene is non-elevated and tunnel; the situation where the image is an elevated bridge without occlusion or without occlusion. If there is no elevated bridge in the road scene of all candidate points of the current trajectory point, it is considered that the attributes of all non-tunnel roads are consistent with the road scene of the current image. If there is an elevated bridge, the unobstructed scene of the elevated bridge is consistent with the attributes of all non-tunnel roads, and the unobstructed scene is consistent with the attributes of the elevated road. The score of each candidate trajectory in terms of road scene semantics is equal to the sum of the scores of all candidate points on the candidate trajectory in terms of road scene semantics.

4. The vehicle trajectory map matching method integrating vehicle-mounted image semantics as claimed in claim 1, characterized in that: In step 2, a convolutional neural network is used to identify the number of lanes in the vehicle image. The lanes on the left side of the solid yellow line and the non-motorized vehicle lane on the right side are ignored. Roads without lane markings are assumed to have one lane by default. The vehicle image recognition results and urban road attributes with more than four lanes are manually corrected to four lanes. The semantic quantification method of the number of lanes is as follows: Among them, α∈[0,1] is the confidence of semantic recognition of vehicle-mounted images, y p ,y r are the number of lanes extracted from the rectified vehicle image and the number of lanes recorded in the urban road attributes, respectively. max and abs are functions for calculating the maximum value and absolute value, respectively. The score of each candidate trajectory in terms of the lane number semantics is equal to the sum of the lane number semantics scores of all candidate points on the candidate trajectory.

5. The vehicle trajectory map matching method integrating vehicle-mounted image semantics as claimed in claim 1, characterized in that: In step 2, a convolutional neural network is used to identify the guide lines that appear in the vehicle image. The identification content includes four aspects: no guide line, guide line appears on the left side of the vehicle, guide line appears on the right side of the vehicle, and guide line appears on both sides of the vehicle. The quantification method of the guide line semantics is as follows: Wherein, α∈[0,1] is the confidence of semantic recognition of vehicle-mounted images. If there is no guide line in the current vehicle-mounted image, the score of all candidate points at the current trajectory point is α; if a guide line appears in the vehicle-mounted image, each candidate point of the current trajectory point is taken as the center, and 5 consecutive candidate points are selected on each candidate trajectory to form a trajectory chain. The overlapping trajectory chains are grouped using a union-find data structure; after grouping, if there is only one trajectory chain in the group, the score of the candidate trajectory at the current trajectory point is 0.8α; for scenes where there are multiple trajectory chains in the group, the trajectory chains are sorted according to the position where the guide line appears, that is, if the current guide line appears on the left side of the vehicle, it proves that the vehicle is driving on the right side, and the score of the trajectory on the right in the same group is α, and the other trajectories decrease by 0.2a from right to left; if the guide line appears on both sides of the vehicle, the score of one or two candidate trajectories in the middle is α, and the other trajectories decrease by 0.2a from the center to both sides. In formula (13), c j represents the number of trajectory chains in the ,th group, and i represents the order of the trajectory chains in the group.

6. The vehicle trajectory map matching method integrating vehicle-mounted image semantics as claimed in claim 1, characterized in that: The specific implementation of step 3 includes the following sub-steps: Step 3.1, data normalization: where x i,j represents the semantic score of the jth vehicle image of the i-th candidate track, min i=1...n x i,j and max i=1...n x i,j They represent the maximum and minimum values ​​of a certain type of vehicle image semantics in all candidate trajectories respectively; Step 3.2, weight calculation, calculate the weight of each type of vehicle image semantics in each candidate track in the corresponding indicators of all candidate tracks: Step 3.3, entropy calculation, calculate the entropy value of the semantic score of each type of vehicle image: Step 3.4, information entropy redundancy calculation, calculate the information entropy redundancy of each type of vehicle image semantics: d j =1-e j (17) Step 3.5, weight calculation, calculate the weight of each vehicle image semantics: Step 3.6, calculate the score of each candidate trajectory, and select the candidate trajectory with the highest overall score as the result of map matching: