CN105957342B - Track grade road plotting method and system based on crowdsourcing space-time big data - Google Patents

Abstract

The present invention provides a lane-level road mapping method based on crowdsourcing spatio-temporal big data, which includes establishing a similarity evaluation model of trajectory vectors, optimizing trajectory based on the growth clustering method of fusion experience knowledge, constructing a Gaussian constraint mixture model, and using The EM algorithm solves the model parameters; detects the lane information, and obtains the initial detection result of the number of lanes in the road section; based on the road construction rules, the initial detection result is corrected; according to the corrected number of lanes, the centerline of the lane is corrected according to the adjacent conditions. The invention reduces the cost of acquiring urban fine road information, and the detection method is simple and easy to implement.

Description

Lane-level road mapping method and system based on crowdsourcing spatio-temporal big data

technical field

The invention relates to high-precision lane-level road mapping of crowdsourced spatiotemporal big data, and belongs to the field of geographic information system and intelligent transportation research.

Background technique

High-precision road maps are the blood and soul of future autonomous driving, and lane-level fine road information is a key component of building high-precision road maps. At present, researchers propose to extract lane lines from high-resolution remote sensing images based on vision methods, or use high-precision laser point cloud data to obtain road detail information, and extract road pavement from a large number of high-precision GPS trajectory data collected by surveying vehicles. Information (road boundaries, number of lanes, lane centerlines). Rogers et al. (1999) was one of the first researchers who attempted to extract road centerlines and lane boundaries using spatio-temporal DGPS (Differential Global Positioning System) trajectory data. Subsequently, on the basis of the research of Rogers et al., the use of spatio-temporal GPS trajectory data to obtain road information gradually developed into an end-to-end model. This end-to-end mode of road information acquisition can be summarized as the following processes: first, optimize the DGPS trajectory data, then match the DGPS trajectory data with the existing map data, fit the spline curve to the road centerline, and finally pass Class methods extract lane information and refine intersection geometry. John Krumm proposed a road information acquisition mode that is separated from the original map. This mode first uses trajectory classification and fusion methods to extract road level information from a large number of DGPS trajectory data, and then uses Gaussian mixture model to extract from a large number of trajectory data belonging to each road segment. lane information. However, these approaches to obtain lane-level fine road information have disadvantages such as high data collection cost, long collection time, slow update speed, and complex data processing.

With the rapid development of sensor technology, wireless communication and network technology, "everyone is a sensor", people's travel will generate a large amount of spatio-temporal trajectory big data, which contains rich fine road information and human behavior information. The collection of trajectory data has gradually evolved from the collection of measuring vehicles by professional departments or by professionals to the form of free and voluntary recording of travel trajectories by non-professionals, and the collection of data has begun to transform into a crowdsourcing model. The vehicle trajectory data (crowdsourcing big data) under the crowdsourcing mode is undoubtedly the best data source that can provide lane-level road information extraction at present. Compared with the existing taxi data, the vehicle trajectory data collected by the crowdsourcing mode belongs to big data (big data refers to the collection of data that cannot be captured, managed and processed by conventional software tools within an affordable time frame, and is A new processing model is needed to have stronger decision-making power, insight discovery power and process optimization ability to adapt to massive, high growth rate and diversified information assets). At present, domestic scholars Tang Luliang et al. (2015, 2016) proposed to use low-precision GPS trajectory data to extract urban lane-level road information, including the number of lanes, lane turning, and lane centerline. Road mapping is a difficult problem faced by scientists all over the world.

Contents of the invention

On the basis of the above research, the present invention proposes a new technical solution for high-precision lane-level road mapping (high-quality trajectory data filtering and high-precision road information extraction) based on crowdsourcing spatiotemporal big data.

The technical solution of the present invention provides a lane-level road mapping method based on crowdsourcing spatio-temporal big data, comprising the following steps,

Step 1, establishing a similarity evaluation model of trajectory vectors, assuming that v _a and v _b are two different trajectory vectors, the similarity evaluation model is as follows,

in, Represents the similarity value between vectors, e is the natural base, ω ₁ and ω ₂ represent the weight values of distance factor diff _Hd and angle factor diff _θab respectively, and ω ₁ +ω ₂ =1; distance factor diff _Hd and angle factor diff _θab represents the distance difference and angle difference of vectors v _a and v _b respectively;

Step 2. Optimizing the trajectory based on the growth clustering method based on the fusion of empirical knowledge, including determining the weight values ω ₁ and ω ₂ of the similarity evaluation model based on the existing high-precision GPS trajectory data and the synchronized low-precision GPS trajectory data, and extracting The prior knowledge of trajectory optimization, based on the similarity between crowdsourcing trajectory data, adopts the growth clustering method to optimize data;

Step 3, build a Gaussian constrained mixture model, and use the EM algorithm to solve the model parameters; the Gaussian constrained mixture model is defined as follows,

Among them, p(x) represents the comprehensive probability value of the Gaussian constrained mixture model, and x represents the sample value to be calculated. When calculating the lane, x represents the ordinate value of the vertical projection of the trajectory point in the judgment window on its longitudinal section; k is the number of Gaussian components, and each Gaussian component corresponds to a lane; ω _j is the weight of the jth Gaussian component, corresponding to the traffic flow of the lane; the parameter μ ₁ ...μ _k is the average value of the trajectory in each Gaussian component, equal to each The center line of each lane, μ _j represents any value in μ ₁ ...μ _k parameters, j=1,2,...,k; σ is the standard deviation of the trajectory in each Gaussian component;

The method for obtaining the quantity k of Gaussian components in the Gaussian constrained mixture model is to calculate the value of the structural risk model, and determine k based on the principle that the value of the structural risk model is the smallest;

Step 4, according to the result obtained in step 3, detect the lane information, and obtain the initial detection result of the number of lanes in the road section; the implementation method is as follows,

Taking all the trajectories on the same road segment as an extraction unit, set a set of trajectories A ^T from the intersection Intersection ₁ to the intersection Intersection ₂ , and start from one end of the trajectory set A ^T to construct a moving rectangular window, where The long side of the moving rectangular window is parallel to the center line of all tracks currently covered, the wide side of the moving rectangular window is perpendicular to the center line of all tracks currently covered, and the center line of the long side of the rectangular window is perpendicular to the center of all track data it covers Line, the center line of the wide side of the rectangular window coincides with the center line of the covered trajectory data;

Start the moving rectangular window from one end of the trajectory set ^AT , and start to translate according to the length of the long side of the rectangular window, and then use the Gaussian constraint mixture model to detect the number of lanes and lane centerlines of the road sections covered in each rectangular window, including moving Rectangular window, project all trajectory points in the moving rectangular window onto the center line of the long side of the rectangular window, and obtain the projected trajectory data set X=(x ₁ ,x ₂ ,…,x _N ), t=1,2 ,3,...,N, where x _t represents the ordinate value of the tth trajectory point after projection, and N is the number of trajectory points participating in the projection; Substitute the trajectory data set X into the Gaussian constraint mixture model to extract the road section in the rectangular window The number of lanes and the centerline of the lanes; assuming the trajectory set A ^T from intersection Intersection ₁ to intersection Intersection ₂ , the rectangular window has been translated l times in total, and the number of lanes determined by each translation is recorded as Nlane _f , f=1, 2,...,l, as the initial detection results of the number of lanes in the road segment;

Step 5, according to the initial detection result of the number of lanes in the road segment obtained in step 4, and based on the road construction rules, the initial detection result is corrected;

In step 6, according to the number of corrected lanes obtained in step 5, the centerline of the lane is corrected according to the adjacent conditions.

Moreover, in step 5, the initial detection result is corrected, and the implementation method is as follows,

The first step, for the number of lanes Nlane _f determined by a translation in step 4, compare Nlane _f+1 with Nlane _f and Nlane _f+2 , if Nlane _f and Nlane _f+1 are different, replace Nlane with Nlane _{f f+1} , f=1,2,...,l-2;

The second step is to classify the results of the first step according to the value and distribution of Nlane _f , assuming there are s classes, recorded as C _g =<Nl _g , nc _g >, Nl _g is the number of lanes of the cluster C _g , nc _g is the total number of C _g , g=1,2,...,s in Nlane _g ;

The third step is to compare C _g+1 and C _g , if Nl _g+1 is different from Nl _g , and nc _g+1 < cv, let Nl _{g of C g replace} Nl g _{+1 of C g +1} , g= 1,2,...,s, complete the final optimization of the number of lanes, where cv is the preset threshold.

Moreover, step 6 corrects the centerline of the lane, and the implementation method is as follows,

Assuming that the number of lanes of a section La is corrected, if the adjacent sections Lb and Lc of La have the same number of lanes as La at the same time, and the number of lanes before and after correction has not changed, then the Lb and Lc The lane centerline is connected to get the final lane centerline of La; if the number of lanes of the adjacent section Lb or Lc of La also changes before and after correction, then the calculation is based on the La correction based on the position of the lane centerline extracted before La correction. Calculate the road centerline of the La section from the position of the front lane centerline, and re-determine the La-corrected lane centerline position according to the La-corrected lane width and the number of lanes.

The present invention provides a lane-level road mapping system based on crowdsourcing spatiotemporal big data, which includes the following modules,

The first module is used to establish a similarity evaluation model of trajectory vectors, assuming v _a and v _b are two different trajectory vectors, and the similarity evaluation model is as follows,

in, Represents the similarity value between vectors, e is the natural base, ω ₁ and ω ₂ represent the weight values of distance factor diff _Hd and angle factor diff _θab respectively, and ω ₁ +ω ₂ =1; distance factor diff _Hd and angle factor diff _θab represents the distance difference and angle difference of vectors v _a and v _b respectively;

The second module is used to optimize the trajectory based on the growth clustering method based on the fusion of empirical knowledge, including determining the weight values ω ₁ and ω of the similarity evaluation model based on the existing high-precision GPS trajectory data and the synchronized low-precision GPS trajectory data _2. Extract the prior knowledge of trajectory optimization, and use the growth clustering method to optimize data based on the similarity between crowdsourced trajectory data;

The third module is used to construct the Gaussian constraint mixture model, and use the EM algorithm to solve the model parameters; the Gaussian constraint mixture model is defined as follows,

Among them, p(x) represents the comprehensive probability value of the Gaussian constrained mixture model, and x represents the sample value to be calculated. When calculating the lane, x represents the ordinate value of the vertical projection of the trajectory point in the judgment window on its longitudinal section; k is the number of Gaussian components, and each Gaussian component corresponds to a lane; ω _j is the weight of the jth Gaussian component, corresponding to the traffic flow of the lane; the parameter μ ₁ ...μ _k is the average value of the trajectory in each Gaussian component, equal to each The center line of each lane, μ _j represents any value in μ ₁ ...μ _k parameters, j=1,2,...,k; σ is the standard deviation of the trajectory in each Gaussian component;

The method for obtaining the quantity k of Gaussian components in the Gaussian constrained mixture model is to calculate the value of the structural risk model, and determine k based on the principle that the value of the structural risk model is the smallest;

The fourth module is used to detect lane information according to the results obtained in the third module, and obtain the initial detection result of the number of lanes in road sections; the implementation method is as follows,

Taking all the trajectories on the same road segment as an extraction unit, set a set of trajectories A ^T from the intersection Intersection ₁ to the intersection Intersection ₂ , and start from one end of the trajectory set A ^T to construct a moving rectangular window, where The long side of the moving rectangular window is parallel to the center line of all tracks currently covered, the wide side of the moving rectangular window is perpendicular to the center line of all tracks currently covered, and the center line of the long side of the rectangular window is perpendicular to the center of all track data it covers Line, the center line of the wide side of the rectangular window coincides with the center line of the covered trajectory data;

Start the moving rectangular window from one end of the trajectory set ^AT , and start to translate according to the length of the long side of the rectangular window, and then use the Gaussian constraint mixture model to detect the number of lanes and lane centerlines of the road sections covered in each rectangular window, including moving Rectangular window, project all trajectory points in the moving rectangular window onto the center line of the long side of the rectangular window, and obtain the projected trajectory data set X=(x ₁ ,x ₂ ,…,x _N ), t=1,2 ,3,...,N, where x _t represents the ordinate value of the tth trajectory point after projection, and N is the number of trajectory points participating in the projection; Substitute the trajectory data set X into the Gaussian constraint mixture model to extract the road section in the rectangular window The number of lanes and the centerline of the lanes; assuming the trajectory set A ^T from intersection Intersection ₁ to intersection Intersection ₂ , the rectangular window has been translated l times in total, and the number of lanes determined by each translation is recorded as Nlane _f , f=1, 2,...,l, as the initial detection results of the number of lanes in the road segment;

The fifth module is used to correct the initial detection results based on the road construction rules according to the initial detection results of the number of lanes in the road section obtained by the fourth module;

The sixth module is used to correct the centerline of the lane according to the adjacent situation according to the corrected number of lanes obtained by the fifth module.

Moreover, in the fifth module, the initial detection result is corrected, and the implementation method is as follows,

In the first step, for the number of lanes Nlane _f determined by a translation in the fourth module, compare Nlane _f+1 with Nlane _f and Nlane _f+2 , if Nlane _f and Nlane _f+1 are different, replace with Nlane _f Nlane _f+1 , f=1,2,...,l-2;

The second step is to classify the results of the first step according to the value and distribution of Nlane _f , assuming there are s classes, recorded as C _g =<Nl _g , nc _g >, Nl _g is the number of lanes of the cluster C _g , nc _g is the total number of C _g , g=1,2,...,s in Nlane _g ;

The third step is to compare C _g+1 and C _g , if Nl _g+1 is different from Nl _g , and nc _g+1 < cv, let Nl _{g of C g replace} Nl g _{+1 of C g +1} , g= 1,2,...,s, complete the final optimization of the number of lanes, where cv is the preset threshold.

Moreover, the sixth module corrects the centerline of the lane, and the implementation method is as follows,

Assuming that the number of lanes of a section La is corrected, if the adjacent sections Lb and Lc of La have the same number of lanes as La at the same time, and the number of lanes before and after correction has not changed, then the Lb and Lc The lane centerline is connected to get the final lane centerline of La; if the number of lanes of the adjacent section Lb or Lc of La also changes before and after correction, then the calculation is based on the La correction based on the position of the lane centerline extracted before La correction. Calculate the road centerline of the La section from the position of the front lane centerline, and re-determine the La-corrected lane centerline position according to the La-corrected lane width and the number of lanes.

The invention constructs a high-precision lane-level road mapping technical solution for crowdsourcing spatio-temporal big data, which reduces the cost of obtaining urban fine road information, and the detection method is simple and easy to implement. The technical solution provided by the present invention includes: firstly, by comparing the spatial similarity between the high-precision vehicle-mounted GPS trajectory and its synchronous low-precision GPS trajectory data, and using the growth clustering method based on empirical knowledge, the positioning accuracy is selected from the crowdsourced trajectory data. Relatively high data; secondly, construct a moving window perpendicular to the trajectory data; then, use the optimized Gaussian mixture model method to conduct longitudinal detection on all trajectories in the road section, and obtain the number of lanes in the detection window; further, use the road Construction rules, that is, the same road section will only add lanes near the intersection, and the number of lanes in the middle of the road section usually remains unchanged. An optimization strategy for the number of lanes is proposed to correct the information on the number of lanes; finally, use the correction The final lane number information is corrected for the extracted lane centerline, and the extraction of lane-level road information of the corresponding road section is completed. The correct rate of judging the number of lanes obtained by the invention is 85%, and the positioning accuracy of the center line of the lane is about 0.35m, which reduces the cost of acquiring urban fine road information, and the detection method is simple and easy to implement.

Description of drawings:

Fig. 1 is the method flowchart of the embodiment of the present invention;

Fig. 2 is a schematic diagram of trajectory vector similarity in an embodiment of the present invention;

Fig. 3 is a schematic diagram of growth clustering based on empirical knowledge according to an embodiment of the present invention;

Fig. 4 is a schematic diagram of the trajectory optimization result of the growth clustering method based on empirical knowledge according to an embodiment of the present invention, wherein Fig. 4a is a schematic diagram of the experimental area of crowdsourcing trajectory data, and Fig. 4b is a schematic diagram of the trajectory optimization result;

Fig. 5 is a schematic diagram of the Gaussian mixture model and the position detection of the lane centerline according to the embodiment of the present invention, wherein Fig. 5a is the optimization result of the Gaussian mixture model, and Fig. 5b is the detection result of the lane centerline position;

6 is a schematic diagram of building a rectangular window to detect lane information according to an embodiment of the present invention;

Fig. 7 is a schematic diagram of optimization of the number of lanes in an embodiment of the present invention;

Fig. 8 is a schematic diagram of the detection result of the number of lanes according to the embodiment of the present invention;

Fig. 9 is a schematic diagram of the optimization result of the number of lanes according to the embodiment of the present invention.

Detailed ways

The technical solution of the present invention will be described in detail below in conjunction with the drawings and embodiments.

The present invention provides a high-precision lane-level road mapping method for crowdsourcing spatiotemporal big data, referring to Fig. 1, the embodiment includes the following steps:

Step 1. Establish a similarity evaluation model of trajectory vectors. According to the driving characteristics of the vehicle, the general driver will drive along the centerline of the lane according to the driving rules. Therefore, without considering the short-term lane-changing behavior, the high-precision trajectory data that can truly describe the driver's driving trajectory is usually concentrated near the centerline of the lane, and the distance between adjacent trajectories is smaller than the lane width. The included angle tends to be around 0°. In order to evaluate the similarity between the trajectories belonging to the center line of the same lane, the present invention establishes a new trajectory vector similarity evaluation model. In terms of similarity measurement, the trajectory vector refers to the unit trajectory vector composed of each trajectory point and its heading angle in a trajectory, that is, the modules of these trajectory vectors are the same and can be defined arbitrarily. The difference between two trajectory vectors is measured in terms of both direction and distance. As shown in Figure 2, N represents the true north direction, v _a <(x _a ,y _a ),(x _a+1 ,y _a+1 )> and v _b <(x _b ,y _b ),(x _{b +1} ,y _b+1 )> are two different trajectory vectors, the azimuth angles of vectors v _a and v _b are θ _a and θ _b respectively, and the similarity evaluation model is as follows:

in, represents the similarity value between vectors, and When the similarity value is 1, it means that the two trajectory vectors are exactly the same, and when the similarity value is 0, it means that the two trajectory vectors are completely dissimilar; e is the natural base; ω ₁ and ω ₂ represent the distance factor (diff _Hd ) and angle respectively The weight value of the factor (diff _θab ), and ω ₁ +ω ₂ =1; diff _Hd and diff _θab represent the distance difference and angle difference of vectors v _a and v _b respectively. Equation 8 defines the distance difference diff _Hd of two trajectory vectors, and Equation 9 defines the angle difference diff _θab of two trajectory vectors:

Dis _confine in Equation 8 is determined by the width of the lane and is a constant used to constrain the similarity of GPS trajectories close to the centerline of each lane on the same lane. Hd _ab is the vertical distance from the starting point of vector v _a to the starting point of v _b , the calculation formula is shown in formula 11; Hd _ba is the vertical distance from the starting point of vector v _b to the starting point of vector v _a , the calculation formula is shown in formula 12; △θ is the vector The heading angle difference between v _a and vector v _b , where the azimuth angles of vector v _a and v _b are θ _a and θ _b respectively, the calculation formulas are as follows:

Δθ＝|θ _a -θ _b | Formula 10

Step 2, trajectory optimization based on the growth clustering method fused with empirical knowledge. The similarity value between trajectory vectors of high-precision trajectory points is also relatively high. When performing trajectory optimization, the existing high-precision GPS trajectory data must be synchronized with the low-precision GPS trajectory data to extract prior knowledge of trajectory optimization. In the embodiment, the accuracy of the high-precision DGPS trajectory and the synchronous low-precision trajectory are 0.5m and 10-15m respectively, and the sampling frequency is 1s. Calculate the similarity between the low-precision trajectory and the DGPS trajectory according to the similarity estimation model established in step 1, and determine the weight value of the similarity evaluation model and acquire empirical knowledge.

1) Determine the weight value of the similarity evaluation model

Similarity is not only used to extract prior knowledge from DGPS and low-precision GPS data, but also used for crowd source data clustering and optimization, so the present invention further proposes to use the correlation between vertical distance, angle difference and measurement error to estimate ω ₁ and the value of _ω2 .

For the GPS track set T=<Trace ₁ ,Trace ₂ ,...,Trace _s >, which contains s GPS tracks: Trace ₁ ,Trace ₂ ,...,Trace _s , the synchronous high-precision DGPS track of the track set T is expressed as DT= <Dt ₁ , Dt ₂ , . . . , Dt _s >, Dt ₁ , Dt ₂ , . . . , Dt _s are trajectory data within the set DT. The accuracies of trajectory set DT and trajectory set T are 0.5m and 10-15m, respectively. Assume that the i-th GPS track Trace _i =<p ₁ , p ₂ ,..., p _n >, p ₁ , p ₂ ,..., p _n represent the track points of the track Trace _i respectively, and there are n track points in total; Dt _i = <rp ₁ ,rp ₂ ,…,rp _n >, rp ₁ ,rp ₂ ,…,rp _n are the track points of the synchronous high-precision track Dt _i of the track Trace _i ; Trace _i ∈ T, Dt _i ∈ DT,i =1,2,...,s. The trajectory vector formed by the trajectory Trace _i and the trajectory points in its synchronous high-precision trajectory Dt _i is expressed as: Tv _i =<v ₁ ,v ₂ ,...,v _n-1 >, v ₁ ,v ₂ ,..., v _n-1 is respectively represented as a trajectory vector composed of the trajectory points p ₁ and p ₂ ,...,p _n-1 and p _n of the trajectory Trace _i ; Dv _i ＝<rv ₁ ,rv ₂ ,...,rv _n-1 >, rv ₁ , rv ₂ ,...,rv _n-1 are respectively represented as trajectory vectors composed of the synchronous high-precision trajectory Dt _i trajectory points rp ₁ and rp ₂ ,...,rp _n-1 and rp _n of the trajectory Trace _i . The distance and angle of Tv _i and Dv _i are respectively expressed as: D _i =<d ₁ ,d ₂ ,…,d _n-1 >,d ₁ ,d ₂ ,…,d _n-1 represents the trajectory vector set Tv _i and The distance from the corresponding trajectory vector v ₁ , v ₂ ,…,v _n-1 to rv ₁ ,rv ₂ ,…,rv _n-1 in Dv _i ; A _i =<a ₁ ,a ₂ ,…,a _{n -1} >,a ₁ ,a ₂ ,…,a _n-1 means the corresponding trajectory vector v ₁ ,v ₂ ,…,v _n-1 and rv ₁ ,rv ₂ , in the trajectory vector set Tv _i and Dv _i ..., rv _n-1 angle difference, i = 1, 2, ..., s. The positioning error of all the trajectory data in the trajectory set T can be expressed as E _i =<ε ₁ ,ε ₂ ,…,ε _n >, where E _i represents the synchronous height of all the trajectory points in the trajectory Trace _i in the trajectory set T and their corresponding The spatial distance of all track points in the precision track Dt _i , ε _j is any error value in the set E _i , where ε _j =|p _j -rp _j |, p _j is any track point in Trace _i , rp _j is the high-precision trajectory point corresponding to p _j , i=1,2,...,s, j=1,2,...,n. In the similarity evaluation model, the calculation formulas of weights ω ₁ and ω ₂ are as follows (r _Dε and r _Aε represent the correlation coefficients of D _i and E _i respectively, and the values of A _i and E _i can be estimated based on the covariance matrix):

ω ₂ =1-ω ₁ Formula 14

2) Prior Knowledge Extraction

Step 1: According to the vector similarity evaluation model proposed in step 1 and the determination of weight values, calculate the similarity between high-precision DGPS and its synchronous low-precision GPS, and compare the location of low-precision GPS trajectory data with its synchronous high-precision The position of DGPS data, calculate the measurement error of low-precision GPS data;

In the second step, according to the similarity obtained in the first step and its corresponding GPS measurement error, an attribute description pair belonging to the same GPS trajectory data is constructed, that is, (similarity value, GPS measurement error) is an attribute of a certain trajectory Description right.

The third step is to select all GPS data whose similarity value is greater than 0.5, greater than 0.6, greater than 0.7..., greater than 0.9 according to the similarity threshold, that is, starting from the similarity value of 0.5, from the attribute description of the GPS data. And count the GPS measurement errors of all GPS data belonging to each similarity threshold, and calculate the average value of these GPS measurement errors and the ratio of the number of data meeting these thresholds to the overall data.

In the fourth step, the similarity threshold set in the third step (such as: the similarity threshold is greater than 0.5, and the similarity threshold is greater than 0.6...) is defined as: Ts _h , h=1,2,3,4,5; The average value of GPS measurement errors meeting these thresholds is defined as Ts _h , h=1,2,3,4,5; the ratio of GPS measurement errors meeting the thresholds to the overall data is defined as: Per _h , h=1,2 ,3,4,5; complete prior knowledge acquisition.

Among them, ST _h represents the data set that satisfies Ts _h , ST _h ∈ T, T is the overall data set of experimental data used for empirical extraction, and the calculation formula of the percentage of ST _h is as follows:

Per _h represents the percentage of ST _h , N(ST _h ) and N(T) are the number of ST _h and T trajectory points respectively, and the calculation formula of ST _h is:

Tε _h is the measurement error of Ts _h , ∑ε is the sum of all track point errors in ST _h , h=1, 2,...,5, Ts _h , Tε _h , Per _h is recorded as RST _h ＝<Ts _h , Tε _h , Per _h >, RST _h is the prior knowledge set, which is the prior knowledge of the growth clustering method.

According to the extracted prior knowledge, the distance weight and angle weight in the similarity evaluation model are calculated, and then based on the similarity between crowdsourcing trajectory data, the growth clustering method is used for data optimization, in which the prior knowledge Tε is the threshold of clustering, Per is used to select the proportion of high-precision data from the entire trajectory cluster, see Figure 3 (v _s in Figure 3 represents the seed vector, and v _sn represents any one of the similarity calculations with the seed vector Vector, Cluster ₁ , Cluster ₂ , and Cluster ₃ represent several trajectory vector clusters obtained after clustering; 'Selected data' in part (d) of Figure 3 represents the final selected trajectory data), growth clustering The main steps of the method are as follows:

Step 1: Initialize all trajectory vectors and mark them as unclustered; initialize the number of the current cluster (CC _L ), record _{C L} =1;

Step 2: If there are unclustered trajectory vectors, randomly select a trajectory vector from the remaining unclustered trajectory vectors as the seed trajectory vector v _s , and the clustering label of the seed trajectory vector is CL(v _s ), And CL(v _s )= _{C L} , as shown in part (a) of Figure 3, enter the third step; if all trajectories have been clustered, as shown in Figure 3 (c), then enter the fifth step;

Step 3: Search for the trajectory vector close to v _s , denoted as v _sn : If Sim(v _s ,v _sn )>Tε is satisfied, v _s and v _sn are fused into a trajectory cluster, and the clustering mark of v _sn is denoted as CL(v _sn ), and CL(v _sn )=CC _L , enter the fourth step; if no trajectory vector meeting the similarity threshold with the current seed vector v _s is found, set CC _L =CC _L +1, return The second step; (Sim(v _s ,v _sn ) is the similarity value of v _s and v _sn , Tε is the similarity threshold)

The fourth step: make the trajectory vector v _sn as the seed trajectory v _s , and return to the third step, as shown in part (b) of Figure 3;

Step 5: Calculate the ratio of the number of trajectory points in all clusters to the total number of trajectory points participating in the clustering based on the results of the first to fourth steps, that is, the final clustering cluster, and then calculate all trajectory clusters according to The ratio of the number of trajectory points to the total number of trajectory points is arranged in reverse order, where Per represents the optimal selectivity of the data. From the proportion of trajectory points of the first cluster, the cumulative summation is performed until Per is satisfied. These clusters participating in the cumulative summation are regarded as The optimal selection of high-precision data, as shown in part (d) of Figure 3, the crowdsourced trajectory data of the selected experimental area is shown in Figure 4a, and the result of trajectory optimization is shown in Figure 4b.

Step 3, build a Gaussian constrained mixture model, and use the EM algorithm to solve the model parameters. On the basis of the existing Gaussian constrained hybrid algorithm for obtaining lane information, the present invention optimizes it. A Gaussian constrained mixture model is defined as follows:

Among them, p(x) represents the comprehensive probability value of the Gaussian constrained mixture model, and x represents the sample value to be calculated (when performing lane calculation, x represents the ordinate value of the vertical projection of the trajectory point in the judgment window on its longitudinal section); k is the number of Gaussian components, that is, the number of Gaussian peaks in the constrained Gaussian mixture model, which represents the number of lanes, and each Gaussian component corresponds to a lane; ω ₁ ... ω _k is the weight of each component, corresponding to the weight of each lane Traffic flow, where the weight value is positive and normalized, that is, ω _j is the weight of the jth Gaussian component, ω _j >0,j=1,2,...,k,ω ₁ +ω ₂ +...ω _k =1 ; The parameter μ ₁ ... μ _k is the average value of the trajectory in each Gaussian component, which is equal to the centerline of each lane, μ _j represents any value in the parameters of μ ₁ ... μ _k , j = 1, 2, ..., k ;σ is the standard deviation of the trajectory in each Gaussian component, and since the width of each lane is usually the same as that of adjacent lanes, σ is set as a constant, usually set to 1.75 (according to domestic road construction standards, the lane width is generally About 3.75m, in the specific implementation process, those skilled in the art can reset according to the road construction standards of the selected area). Use the EM method to solve the model parameters: θ _j ^(m) (ω _j ^(m) , μ _j ^(m) ,σ ^(m) ), j=1,2,...,k, where m is the number of iterations. At present, there are many mature methods on how to use the EM algorithm to solve the parameters of the Gaussian mixture model. In the specific implementation process, technical personnel can refer to the existing methods, and the present invention will not repeat them.

The key to the Gaussian constrained mixture model is to obtain the number of Gaussian components, that is, to calculate the value of the structural risk model corresponding to each value of k, and then select the k corresponding to the minimum value of the structural risk model as the number of lanes. The construction method of the structural risk model is as follows:

k=min(R _srm (p(x _i |θ _k ))) Formula 19

L(x _i ,p(x _i |θ _k ))=-log(p(x _i |θ _k )) Formula 20

R _srm (p( _xi |θ _k )) in Equation 2 is a structural risk model, L( _xi ,p( _xi |θ _k )) is an empirical risk model for assessing fitness, J(p( x _i |θ _k )) is a regular term, which is used to mark the complexity of the model, which is expressed as: J _TSW (p(x _i |θ _k )), λ>0 is a regular parameter, p(x _i |θ _k ) represents the Gaussian probability value of the sample value x _i under the condition of the model parameter θ _k , where the model parameter θ _k can be expressed as: θ _k (ω _k ,μ _k ,σ), n represents the number of samples, i=1, 2,...,n; the calculation formula is as follows:

Among them, _Dw is the tile width of the optimized trajectory on the road surface, as shown in Figure 5, the '1 ^st Gaussian component' marked by the long dashed line in Figure 5a represents the first component of the Gaussian mixture model, and the ' ^2nd Gaussian component' marked by the short dashed line Gaussiancomponent' represents the second component of the Gaussian mixture model; the parameters μ ₁ and μ ₂ in Figure 5b correspond to the mean value of the first Gaussian component and the mean value of the second Gaussian component in Figure 5a respectively, that is, The lane centerline location of the first lane and the lane centerline location of the second lane. (how to obtain the tiling width of the track on the road surface has many methods to propose at present, in the specific implementation process, those skilled in the art can choose by themselves, the present invention will not repeat them), wherein k represents the number of possible lanes, according to the current domestic According to road construction standards, the number of urban lanes generally includes two lanes, three lanes, four lanes, and five lanes, that is, k=2,3,4,5. Δμ ^k is the change value of the average value μ _j of two adjacent Gaussian peaks in the Gaussian component corresponding to k, and its change value also reflects the detection lane width, j=1,2,...,k. The calculation method of Δμ ^k is shown in Formula 6. In Formula 6, κ, η, and γ _ij are the hyperparameters of the EM algorithm based on maximum a posteriori estimation; x represents the sample value (when performing lane calculation, x represents the trajectory point in the judgment window ordinate value of the vertical projection on its longitudinal section); n represents the number of sample data, ω _j+1 represents the weight value of the j+1th Gaussian component, j=1,2,...,k-1, k is Number of Gaussian components, k=2, 3, 4, 5; there are very mature parameter recommendations in the existing research, and the specific implementation can refer to the values given in the existing method cases for calculation, and the details will not be repeated.

Step 4. According to the optimized Gaussian mixture model method described in step 3, complete high-precision lane-level road mapping, realize detection of lane information, and obtain the initial detection result of the number of lanes in the road section. In the process of implementing high-precision lane-level road mapping, use existing methods to use all trajectories on the same road segment as an extraction unit (specifically, how to classify all trajectories on the same road segment into an extraction unit, currently there are There are many mature methods, those skilled in the art can refer to the existing methods during the specific implementation process, and the present invention will not repeat them).

Assume that a set of trajectory sets A ^T from intersection Intersection ₁ to intersection Intersection ₂ is given, and a moving rectangular window is constructed from one end of the trajectory set A ^T , as shown in Figure 6. The length and width of the moving rectangular window are rh and rw respectively (the length and width of the rectangular window are recommended to be defined as 10m and 30m, and those skilled in the art can preset the values during specific implementation), wherein the long side of the moving rectangular window is parallel to The center line of all tracks currently covered, the wide side of the moving rectangular window is perpendicular to the center line of all tracks currently covered, the center line of the long side of the rectangular window is perpendicular to the center line of all track data covered by it, and the center line of the wide side of the rectangular window How to obtain the centerline of a piece of trajectory data coincident with the centerline of the track data covered by it, there are many methods at present, those skilled in the art can refer to the existing methods in the specific implementation process, and the present invention will not repeat them.

The preferred implementation mode further provided by the embodiment is as follows:

The moving rectangular window starts from one end of the trajectory set ^AT , and starts to translate according to the length of the long side of the rectangular window, and uses the Gaussian constraint mixture model to detect the number of lanes and the lane centerline of the road section covered in each rectangular window. The specific method includes: according to the constructed rectangular window, project all the trajectory points in the rectangular window onto the center line of the long side of the rectangular window to obtain the projected trajectory data set X=(x ₁ ,x ₂ ,…,x _N ), t=1, 2, 3, ..., N, where x _t represents the ordinate value of the tth trajectory point after projection, and N is the number of trajectory points participating in the projection. As described in step 3, the trajectory data set X is substituted into the corresponding calculation formula (that is, the Gaussian constrained mixture model shown in formula 17), and the number of lanes and the lane centerline of the road section in the rectangular window are extracted. Assuming that the trajectory set A ^T from the intersection Intersection ₁ to the intersection Intersection ₂ , the rectangular window has been translated l times in total, and the number of lanes determined by each translation is recorded as Nlane _f , f=1,2,...,l, as the road The initial detection result of the number of lanes in the road segment.

Step 5: According to the initial detection result of the number of lanes in the road section obtained in step 4, based on the road construction rules, an optimization strategy for the number of lanes is proposed, and the initial detection result is corrected. As shown in Figure 7, there are road sections Part ₁ , Part ₂ , and Part ₃ , where there are additional lane areas in Part ₁ and Part ₃ . In most cases, the road section between two intersections will always have additional lanes near the intersection, while the number of lanes in the middle of the road section usually remains unchanged. Therefore, the present invention proposes a method for optimizing the extraction results of the number of lanes. method, the specific method is as follows:

Step 1: For the number of lanes Nlane _f determined by a translation in step 4, compare Nlane _f+1 with Nlane _f and Nlane _f+2 , if Nlane _f and Nlane _f+1 are different, replace Nlane with Nlane _{f f+1} , f=1,2,...,l-2.

The second step: Classify the results of the first step according to the value and distribution of Nlane _f , for example, if the values of Nlane _e , Nlane _e+1 , Nlane _e+2 ,...,Nlane _e+c are the same, they are classified into one category, where e<l, e+c<l. Suppose there are s classes, recorded as C _g ＝<Nl _g , nc _g >, Nl _g is the number of lanes of cluster C _g , nc _g is the number of lanes belonging to C _g , g=1,2,...,s in Nlane _g The total number.

Step 3: Compare C _g+1 and C _g , if Nl _g+1 is different from Nl _g , and nc _g+1 < cv (cv is a threshold value that depends on road construction rules, and its modification rules are mainly reflected in urban roads When approaching the intersection, there is a buffer area for adding lanes. For example, according to the current road design regulations, the length of the lane addition is about 50m, that is, the length of the new lane is 50m from the road section at the intersection, so When the division section of the lane quantity judgment is set to 10 meters, the present invention recommends that cv be set to 5, and those skilled in the art can preset the value voluntarily during specific implementation), so that the Nl g of C _g replaces the Nl _g of C _{g +1 +1} , g=1,2,…,s, complete the final optimization of the number of lanes. The detection result of the number of lanes is shown in Figure 8, and the detection optimization of the number of lanes is shown in Figure 9 (the abscissa in Figure 8 represents the number of movements during the sliding detection process of the moving window, and the vertical coordinate represents the detection result of the number of lanes in each sliding detection process).

In step 6, according to the corrected number of lanes obtained in step 5, the corresponding lane centerline is corrected. When the number of lanes in a section of road is corrected, the corresponding centerline of the lane is also corrected using the principle of adjacency. Specific methods include:

Assuming that the number of lanes of a certain road segment La has been corrected, then the adjacent road segments before and after La are searched. If the road segments Lb and Lc adjacent to La have the same number of lanes as La at the same time, and the number of lanes of these road segments before and after correction is not the same If there is a change, then connect Lb with the lane centerline of Lc to obtain the final lane centerline of La; if the number of lanes of Lb or Lc adjacent to La also changes before and after correction, then extract according to La before correction The position of the center line of the lane, the calculation is based on the position of the center line of the lane before the correction of La to calculate the center line of the road in the La section. ) Redefine the position of the centerline of the lane after La correction according to the lane width and the number of lanes after La correction, that is, starting from the road centerline according to the La road section, the number of lanes and the lane width are equidistant parallel to the road centerline to obtain the La correction The lane centerline position of each lane.

Based on the present invention, it is possible to conveniently obtain the lane information of the urban road from the GPS trajectory data, and provide basic road network data for future intelligent navigation and unmanned driving.

During specific implementation, the method provided by the present invention can realize the automatic operation process based on software technology, and can also realize the corresponding system in a modular manner.

The present invention provides a lane-level road mapping system based on crowdsourcing spatiotemporal big data, which includes the following modules,

The first module is used to establish a similarity evaluation model of trajectory vectors, assuming v _a and v _b are two different trajectory vectors, and the similarity evaluation model is as follows,

in, Represents the similarity value between vectors, e is the natural base, ω ₁ and ω ₂ represent the weight values of distance factor diff _Hd and angle factor diff _θab respectively, and ω ₁ +ω ₂ =1; distance factor diff _Hd and angle factor diff _θab represents the distance difference and angle difference of vectors v _a and v _b respectively;

The second module is used to optimize the trajectory based on the growth clustering method based on the fusion of empirical knowledge, including determining the weight values ω ₁ and ω of the similarity evaluation model based on the existing high-precision GPS trajectory data and the synchronized low-precision GPS trajectory data _2. Extract the prior knowledge of trajectory optimization, and use the growth clustering method to optimize data based on the similarity between crowdsourced trajectory data;

The third module is used to construct the Gaussian constraint mixture model, and use the EM algorithm to solve the model parameters; the Gaussian constraint mixture model is defined as follows,

Among them, p(x) represents the comprehensive probability value of the Gaussian constrained mixture model, and x represents the sample value to be calculated. When calculating the lane, x represents the ordinate value of the vertical projection of the trajectory point in the judgment window on its longitudinal section; k is the number of Gaussian components, and each Gaussian component corresponds to a lane; ω _j is the weight of the jth Gaussian component, corresponding to the traffic flow of the lane; the parameter μ ₁ ...μ _k is the average value of the trajectory in each Gaussian component, equal to each The center line of each lane, μ _j represents any value in μ ₁ ...μ _k parameters, j=1,2,...,k; σ is the standard deviation of the trajectory in each Gaussian component;

The method for obtaining the quantity k of Gaussian components in the Gaussian constrained mixture model is to calculate the value of the structural risk model, and determine k based on the principle that the value of the structural risk model is the smallest;

The fourth module is used to detect lane information according to the results obtained in the third module, and obtain the initial detection result of the number of lanes in road sections; the implementation method is as follows,

Taking all the trajectories on the same road segment as an extraction unit, set a set of trajectories A ^T from the intersection Intersection ₁ to the intersection Intersection ₂ , and start from one end of the trajectory set A ^T to construct a moving rectangular window, where The long side of the moving rectangular window is parallel to the center line of all tracks currently covered, the wide side of the moving rectangular window is perpendicular to the center line of all tracks currently covered, and the center line of the long side of the rectangular window is perpendicular to the center of all track data it covers Line, the center line of the wide side of the rectangular window coincides with the center line of the covered trajectory data;

Start the moving rectangular window from one end of the trajectory set ^AT , and start to translate according to the length of the long side of the rectangular window, and then use the Gaussian constraint mixture model to detect the number of lanes and lane centerlines of the road sections covered in each rectangular window, including moving Rectangular window, project all trajectory points in the moving rectangular window onto the center line of the long side of the rectangular window, and obtain the projected trajectory data set X=(x ₁ ,x ₂ ,…,x _N ), t=1,2 ,3,...,N, where x _t represents the ordinate value of the tth trajectory point after projection, and N is the number of trajectory points participating in the projection; Substitute the trajectory data set X into the Gaussian constraint mixture model to extract the road section in the rectangular window The number of lanes and the centerline of the lanes; assuming the trajectory set A ^T from intersection Intersection ₁ to intersection Intersection ₂ , the rectangular window has been translated l times in total, and the number of lanes determined by each translation is recorded as Nlane _f , f=1, 2,...,l, as the initial detection results of the number of lanes in the road segment;

The fifth module is used to correct the initial detection results based on the road construction rules according to the initial detection results of the number of lanes in the road section obtained by the fourth module;

The sixth module is used to correct the centerline of the lane according to the adjacent situation according to the corrected number of lanes obtained by the fifth module.

For the specific implementation of each module, reference may be made to the corresponding steps, which will not be described in detail in the present invention.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or replace them in similar ways, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (6)

1. A lane-level road mapping method based on crowdsourcing space-time big data is characterized in that: comprises the following steps of (a) preparing a solution,

step 1, establishing a similarity evaluation model of a track vector, and setting v _a And v _b Are two different trajectory vectors, the similarity evaluation model is as follows,

wherein,representing the similarity value between vectors, e being a natural base number, omega ₁ And ω ₂ Respectively represent distance factors diff _Hd And angle factor diff _θab And ω is ₁ +ω ₂ =1; distance factor diff _Hd And angle factor diff _θab Respectively represent vectors v _a And v _b The distance difference and the angle difference of (a);

step 2, carrying out track optimization based on a growth clustering method of fusion empirical knowledge, wherein the track optimization comprises the step of determining a weight value omega of a similarity evaluation model according to the existing high-precision GPS track data and the synchronous low-precision GPS track data ₁ And ω ₂ Extracting prior knowledge of track optimization, and optimizing data by adopting a growth clustering mode based on the similarity between crowdsourcing track data;

step 3, constructing a Gaussian constraint mixed model, and solving model parameters by using an EM (effective electromagnetic) algorithm; the gaussian-constrained hybrid model is defined as follows,

wherein p (x) is represented as the comprehensive probability value of the Gaussian constraint mixed model, x represents a sample value to be calculated, and x represents the vertical coordinate value of the vertical projection of the track point in the judgment window on the longitudinal section when the lane calculation is carried out; k is the number of Gaussian components, each corresponding to a lane; omega _j Is the weight of the jth Gaussian component and corresponds to the traffic flow of the lane; parameter mu ₁ …μ _k Is the average of the trajectories in each Gaussian component, equal to the centerline of each lane, μ _j Represents mu ₁ …μ _k Any value within the parameters, j =1,2, \8230;, k; σ is the standard deviation of the trajectories in each Gaussian component;

the number k of Gaussian components in the Gaussian constrained hybrid model is obtained by calculating the value of the structural risk model and determining k on the basis of the minimum value of the structural risk model;

step 4, detecting lane information according to the result obtained in the step 3 to obtain a primary detection result of the number of lanes of the road section; the implementation mode is as follows,

setting a given group of Intersection intersections by taking all tracks on the same road section as an extraction unit ₁ Intersection to Intersection interaction ₂ Track set A of ^T From the set of trajectories A ^T Starting from one end of the window, constructing a moving rectangular window, wherein the long side of the moving rectangular window is parallel to the central lines covering all tracks at present, the wide side of the moving rectangular window is perpendicular to the central lines covering all tracks at present, the central line of the long side of the rectangular window is perpendicular to the central lines of all track data covered by the rectangular window, and the central line of the wide side of the rectangular window is superposed with the central line of the track data covered by the rectangular window;

moving rectangular window from trajectory set A ^T The method comprises the steps of starting translation according to the length of the long edge of the rectangular window, sequentially detecting the number of lanes and the center line of the lanes of a road section covered in each rectangular window by utilizing a Gaussian constraint hybrid model, and projecting all track points in the movable rectangular window to the long edge of the rectangular window according to the movable rectangular windowOn the central line, a projected trajectory data set X = (X) is obtained ₁ ,x ₂ ,…,x _N ) Wherein x is ₁ ,x ₂ ,…,x _N The vertical coordinate values of N track points are shown in 1,2,3, \ 8230after projection, and N is the number of the track points participating in projection; substituting the track data set X into a Gaussian constraint hybrid model, and extracting the number of lanes and the center line of the lanes of a road section in a rectangular window; suppose Intersection interaction ₁ Intersection to Intersection interaction ₂ Track set A of ^T The rectangular window is translated for one time, and the number of lanes determined by each translation is recorded as Nlane _f L, as a primary detection result of the number of lanes of the road section;

step 5, correcting the primary detection result based on the road construction rule according to the primary detection result of the number of lanes of the road section acquired in the step 4;

and 6, correcting the lane center line according to the adjacent condition according to the corrected lane number obtained in the step 5.

2. The method of claim 1, wherein the method comprises: in step 5, the primary detection result is corrected in the following way,

first, the number of lanes Nlane determined for a certain translation in step 4 _f Comparison of Nlane _f+1 And Nlane _f 、Nlane _f+2 If Nlane _f And Nlane _f+1 If different, use Nlane _f Replacement of Nlane _f+1 ，f＝1,2,…,l-2；

Second step, according to Nlane _f The result of the first step is classified by the value and distribution of (C), and s classes are provided and are marked as C _g ＝<Nl _g ,nc _g >,Nl _g Is the g-th cluster C _g Number of lanes of nc _g Number of lanes Nlane determined for each translation _f In (C) _g G =1,2, \ 8230;, total number of s;

thirdly, comparing the g +1 th cluster C _g+1 And the g cluster C _g If the g +1 th clusterC _g+1 Number of lanes Nl _g+1 Is different from Nl _g And the number of lanes Nlane determined for each translation _f Belong to C _g+1 Total number of (nc) _g+1 &lt, cv, order C _g Nl of _g Replacement C _g+1 Nl of _g+1 G =1,2, \8230s, s, the final optimization of the lane number results is done, where cv is a preset threshold.

3. The method of lane-level road mapping based on crowd-sourced spatio-temporal big data of claim 1 or 2, characterized in that: step 6, correcting the center line of the lane in the following way,

if the number of lanes of a certain section of road La is corrected, if the number of lanes of adjacent road sections Lb and Lc of La simultaneously meets the requirement that the number of lanes is the same as that of the lanes of La, and the number of lanes before and after correction does not change, then connecting the lane center lines of Lb and Lc to obtain the final lane center line of La; if the number of the lanes of the La adjacent road sections Lb or Lc is changed before and after the correction, calculating the road center line of the La road section estimated based on the lane center line position before the La correction according to the lane center line position extracted before the La correction, and re-determining the lane center line position after the La correction according to the lane width and the lane number after the La correction.

4. The utility model provides a lane level road mapping system based on crowd-sourced space-time big data which characterized in that: comprises the following modules which are used for realizing the functions of the system,

a first module for establishing a similarity evaluation model of the track vector, and setting v _a And v _b Are two different trajectory vectors, the similarity evaluation model is as follows,

wherein,representing the similarity value between vectors, e being a natural base，ω ₁ And ω ₂ Respectively represent distance factors diff _Hd And angle factor diff _θab And a weight value of ω ₁ +ω ₂ =1; distance factor diff _Hd And angle factor diff _θab Respectively represent vectors v _a And v _b Distance difference and angle difference of (a);

the second module is used for carrying out track optimization based on a growth clustering method of fusion empirical knowledge and comprises the steps of determining a weight value omega of a similarity evaluation model according to existing high-precision GPS track data and synchronous low-precision GPS track data ₁ And ω ₂ Extracting prior knowledge of track optimization, and optimizing data by adopting a growth clustering mode based on the similarity between crowdsourcing track data;

the third module is used for constructing a Gaussian constraint mixed model and solving model parameters by using an EM (effective electromagnetic) algorithm; the gaussian constrained hybrid model is defined as follows,

p (x) is expressed as a comprehensive probability value of the Gaussian constraint hybrid model, x is expressed as a sample value to be calculated, and x represents a vertical coordinate value of a track point in a judgment window vertically projected on a longitudinal section of the track point when lane calculation is carried out; k is the number of gaussian components, each corresponding to a lane; omega _j Is the weight of the jth Gaussian component and corresponds to the traffic flow of the lane; parameter mu ₁ …μ _k Is the average of the trajectories in each Gaussian component, equal to the centerline of each lane, μ _j Represents mu ₁ …μ _k Any value within the parameter, j =1,2, \ 8230;, k; σ is the standard deviation of the trajectory in each Gaussian component;

the number k of Gaussian components in the Gaussian constraint mixed model is obtained by calculating the value of the structural risk model and determining k according to the principle that the value of the structural risk model is minimum;

the fourth module is used for detecting lane information according to the result obtained by the third module to obtain a primary detection result of the number of lanes on the road section; the implementation is as follows, and the method,

setting a given group of Intersection intersections by taking all tracks on the same road section as an extraction unit ₁ Intersection to Intersection interaction ₂ Track set A of ^T From the set of trajectories A ^T Starting from one end of the window, constructing a moving rectangular window, wherein the long side of the moving rectangular window is parallel to the central lines covering all tracks at present, the wide side of the moving rectangular window is perpendicular to the central lines covering all tracks at present, the central line of the long side of the rectangular window is perpendicular to the central lines of all track data covered by the rectangular window, and the central line of the wide side of the rectangular window is superposed with the central line of the track data covered by the rectangular window;

moving rectangular window from trajectory set A ^T The method comprises the steps of sequentially detecting the number of lanes and the lane center line of a road section covered in each rectangular window by using a Gaussian constraint hybrid model according to a moving rectangular window, and projecting all track points in the moving rectangular window onto the long edge center line of the rectangular window to obtain a projected track data set X = (X is the length of the long edge of the rectangular window), wherein the translation is started according to the length of the long edge of the rectangular window ₁ ,x ₂ ,…,x _N ) T =1,2,3, \8230, N, where x ₁ ,x ₂ ,…,x _N The vertical coordinate values of N track points are shown in 1,2,3, \8230afterprojection, and N is the number of the track points participating in projection; substituting the track data set X into a Gaussian constraint hybrid model, and extracting the number of lanes and the center line of the lanes of a road section in a rectangular window; suppose that Intersection interaction is driven from ₁ Intersection to Intersection interaction ₂ Track set A of ^T The rectangular window is translated for one time, and the number of lanes determined by each translation is recorded as Nlane _f L, as a primary detection result of the number of lanes of the road section;

the fifth module is used for correcting the primary detection result based on the road construction rule according to the primary detection result of the number of lanes of the road section acquired by the fourth module;

and the sixth module is used for correcting the lane center line according to the adjacent condition and the corrected lane number obtained by the fifth module.

5. The system of claim 4, wherein the system is configured to map roads at a lane level based on crowd-sourced spatiotemporal big data: the fifth module corrects the primary detection result, and the implementation manner is as follows,

first, the number of lanes Nlane determined for a certain translation in the fourth module _f Comparison of Nlane _f+1 And Nlane _f 、Nlane _f+2 If Nlane _f And Nlane _f+1 If different, use Nlane _f Replacement of Nlane _f+1 ，f＝1,2,…,l-2；

Second step, according to Nlane _f The result of the first step is classified by the value and distribution of (C), and s classes are provided and are marked as C _g ＝<Nl _g ,nc _g >,Nl _g Is the g-th cluster C _g Number of lanes, nc _g Number of lanes Nlane determined for each translation _f In the genus of C _g G =1,2, \ 8230;, total number of s;

thirdly, comparing the g +1 th cluster C _g+1 And the g cluster C _g If the g +1 th cluster C _g+1 Number of lanes Nl _g+1 Is different from Nl _g And the number of lanes Nlane determined for each translation _f Belong to C _g+1 Total number of (nc) _g+1 &lt, cv, order C _g Nl of _g Replacement C _g+1 Nl of _g+1 G =1,2, \8230s, s, the final optimization of the lane number results is done, where cv is a preset threshold.

6. The system of claim 4 or 5, wherein the system is configured to map roads at lane level based on crowd-sourced spatiotemporal big data: the sixth module corrects the center line of the lane in the following way,

if the number of lanes in a certain section of road section La is corrected, if the number of lanes in the adjacent road sections Lb and Lc of La simultaneously meets the requirement that the number of lanes is the same as that of the lanes in La, and the number of lanes before and after correction is not changed, then connecting the lane center lines of Lb and Lc to obtain the final lane center line in La; if the number of the lanes of the La adjacent road sections Lb or Lc is changed before and after the correction, calculating the road center line of the La road section calculated based on the position of the lane center line before the La correction according to the position of the lane center line extracted before the La correction, and re-determining the position of the lane center line after the La correction according to the width and the number of the lanes after the La correction.