CN104504897A - Intersection traffic flow characteristic analysis and vehicle moving prediction method based on trajectory data - Google Patents

Abstract

The invention discloses an intersection traffic flow characteristic analysis and vehicle moving prediction method based on trajectory data, and belongs to the technical field of intelligent traffic system and traffic flow parameter acquisition. The method starts with space transient analysis of a vehicle original trajectory, and describes and analyzes trajectory local geometrical characteristics at different angles, forms a multilevel spectral clustering processing framework based on a vehicle original rough movement track, and automatically extracts and analyzes a plurality of traffic direction modes of an intersection included in the trajectory data. With the basis, the method can acquire intersection sub-phase (signal control intersection) traffic flow and travel time of vehicles in all directions passing through the intersection, and other detailed traffic characteristic parameters, as important complement of conventional traffic data. Through tracking travelling tracks of all moving vehicles at present moment, a traffic direction trajectory mode matching method is used to predict the next behavior of the vehicles, thereby being beneficial for warning safety risks which may exist on an intersection in real time.

Description

A kind of intersection traffic properties of flow based on track data is analyzed and vehicle movement Forecasting Methodology

Technical field

The invention belongs to intelligent transportation system (machine vision and image procossing) and traffic flow parameter acquisition technique field, be specifically related to a kind of intersection traffic properties of flow analysis based on the original rough track data of vehicle and vehicle movement Forecasting Methodology.

Background technology

Intersection is the important component part of urban road system.Observed behavior for intersection traffic stream operation characteristic has important theory and practical significance undoubtedly, it will be intersection capacity, incur loss through delay and service level analysis, intersection channelizing design and traffic organization optimization, and crossing control and management application etc. provides important theoretical foundation (X.Li, X.Li, D.Tang, X.Xu.Deriving features of traffic flow around an intersection from trajectories of vehicles [C] .18th International Conference on Geoinformatics, Beijing, 2010:1-5).Meanwhile, along with the significantly increase of the Current City Road volume of traffic, traffic violation problem becomes increasingly conspicuous, and intersection order is often very chaotic, becomes road traffic accident multiplely.Therefore for crossing Exploration on Train Operation Safety, the research to urban road vehicle motion forecast method must be strengthened, in order to vehicle safety early warning task can be completed further.

Current, along with the growing of transport need and the needs of traffic control, multiple sensors is widely used in traffic condition detection.Compared to the conventional traffic stream acquisition technique of the registration of vehicle in an indirect way such as on-the-spot manual testing and ground induction coil detecting device, video encoder server and watch-dog carry out the mobility of registration of vehicle in a straightforward manner, can record the operational process of the numerous vehicle in crossing in detail and influence each other.The vehicle movement initial trace data gathered by traffic video treatment technology, a beyond doubt important basic data source (Z.Fu, W.Hu, T.Tan.Similarity based vehicle trajectory clustering and anomaly detection [C] .IEEE International Conference on Image Processing.2005, 2:602-605) (X.Li, W.Hu, W.Hu.A coarse-to-fine strategy for vehicle motion trajectory clustering [C] .Proceedings of the 18th International Conference on Pattern Recognition.2006, 1:591-594).For specific traffic environment, tradition method of trajectory clustering hypothesis has existed or can obtain error free and unremitting moving object track (B.T.Morris easily, M.M.Trivedi.A survey of vision-based trajectory learning and analysis for surveillance [J] .IEEE Trans.on Circuits and Systems for Video Technology.2008, 18 (8): 1114-1127) (S.Atev, G.Miller, N.P.Papanikolopoulos.Clustering of vehicle trajectories [J] .IEEE Trans.on Intelligent Transportation Systems.2010, 11 (3): 674-657).Due to the complicacy of traffic environment itself, in the process of process real video stream, the reliability of vehicle detection and track algorithm is relatively low, and this will cause vehicle movement track result to there are a series of serious problems, such as fragment, tracking interruption and error hiding etc.Therefore, people improve track quality often through manual synchronizing.But make to become impossible by manual synchronizing due to following 2 reasons: (1), along with the sharply increase of traffic video data, manual synchronizing expends time in very much, adopts manual synchronizing to ensure that the quality of data will become impossible completely; (2) be difficult to avoid introducing less desirable artificial deviation by manual operation.In sum, current work relies on manual synchronizing very consuming time more, thus be difficult to obtain extensive high-quality intersection vehicles motion trace data, finally cause not possessing the condition of carrying out intersection traffic stream operation characteristic site-test analysis and vehicle movement prediction work.

Summary of the invention

Vehicle movement track is applied to the research of DETECTION OF TRAFFIC PARAMETERS by the present invention, root problem to be solved is when not having manual synchronizing, from the Vehicle tracing data of original coarse (inferior quality), directly robustly find the inherent traffic flow pattern in crossing.Based on the traffic direction pattern that the analysis of video frequency vehicle actual measurement track data clusters obtains, can identify intersection traffic environment and strengthen understanding, the final clear true stroke describing intersection vehicles motor pattern and vehicle pass-through.In order to solve the problem being difficult to obtain extensive high-quality intersection vehicles motion trace data, the present invention extracts based on theory by local robust features, propose a kind of analytical approach (S.Huet adopting analysis track microcosmic geometric properties, E.Karatekin, V.S.Tran, I.Fanget, S.Cribier, J.Henry.Analysis of transient behavior in complex trajectories:application to secretory vesicle dynamics [J] .Biophysical Journal.2006, 91 (9): 3542-3559) (J.A.Helmuth, C.J.Burckhardt, P.Koumoutsakos, U.F.Greber, I.F.Sbalzarini.A novel supervised trajectory segmentation algorithm identifies distinct types of human adenovirus motion in host cells [J] .Journal of Structural Biology.2007, 159 (3): 347-358), the local geometric features of specification and analysis track from different perspectives, directly process for coarse initial trace data, propose a set of general multi-level spectral clustering process framework based on vehicle original coarse movement track, the multiple traffic direction pattern in crossing that automatic extraction and analysis (data mining) track data comprise.

Intersection traffic properties of flow based on track data is analyzed and a vehicle movement Forecasting Methodology, it is characterized in that, comprises the steps:

Step 1: the intersection vehicles based on motion tracking move original rough track obtain, set up crossing extensive vehicle movement track data set;

Step 1.1: intersection vehicles motion initial trace gathers

Adopt the tracking based on image block in OpenCV, automatically extract crossing moving vehicle initial trace data, and be expressed as vehicle movement point sequence T:

T＝{t ₁,t ₂,…,t _i，…，t _n}

＝{(x ₁,y ₁),(x ₂,y ₂),…,(x _i,y _i),…,(x _n,y _n)}

And track step sequence S:

S＝{s ₁,s ₂,…,s _i,…,s _n-1}

＝{(δx ₁,δy ₁),…,(δx _i,δy _i),…,(δx _n-1,δy _n-1)}

Wherein, t _i=(x _i, y _i) represent the position at moving vehicle i-th sampled point center, s _i=(δ x _i, δ y _i)=(x _i+1-x _i, y _i+1-y _i) represent the deviation of neighbouring sample dot center, n represent track of vehicle comprise the sum of sampled point;

Step 1.2: original vehicle movement locus pre-service

Carry out following smoothing processing for every bar track: (1) considers sample noise, as the distance between fruit dot is enough little, just continuous print point combined and replaced by first point; (2) complete deletion length is less than the short track of predefine threshold value; (3) the smoothing process of mean filter is adopted, to retain prototype structure essential characteristic (selecting level and smooth step number w=7 in experiment):

$\stackrel{\‾}{t_i}=\frac{1}{w}\underset{j=i-w/2}{\overset{i+w/2}{\mathrm{\Σ}}}{t}_j$

Step 2: the intersection vehicles motor pattern based on vehicle original rough track Multi-layer Spectral cluster learns

Step 2.1: multi-level track characteristic extracts

The present invention uses linearity and flexibility, trajectory direction histogram and three kinds, center feature to carry out the local feature of Integrative expression track respectively;

Described linearity and flexibility refer to the tolerance that vehicle operating orientation average changes, if α _irepresent s _iwith s _i+1orientation angle change between step, establishes simultaneously and is changed to positive dirction left;

Linearity is defined as follows:

$p_1=\frac{1}{l_w-1}\underset{i=j}{\overset{j+l_w-2}{\mathrm{\Σ}}}\cos {\mathrm{\α}}_j$

Flexibility is defined as follows:

$p_2=\frac{1}{l_w-1}\underset{i=j}{\overset{j+l_w-2}{\mathrm{\Σ}}}\sin {\mathrm{\α}}_j$

Described trajectory direction histogram TDH is a kind of feature representation method of novel description trajectory direction, first, calculates the deflection β of a jth point in track by the following method _j:

${\mathrm{\&beta;}}_j=\left\{\begin{array}{c}\arctan \left(\frac{\mathrm{dy}}{\mathrm{dx}}\right),\mathrm{dx}>0\\ \arctan \left(\frac{\mathrm{dy}}{\mathrm{dx}}\right)+\mathrm{\&pi;},\mathrm{dx}<0,\mathrm{dy}>0\\ \arctan \left(\frac{\mathrm{dy}}{\mathrm{dx}}\right)-\mathrm{\&pi;},\mathrm{dx}<0,\mathrm{dy}<0\end{array}\right.$

Wherein, (dx is required _j) ²+ (dy _j) ²≠ 0 and β _j∈ [-π, + π), then, by interval [-π, + π) be evenly divided into N number of equal-sized director interval, and by all points in identical strip path curve according to the different mappings of deflection in corresponding director interval, finally, carry out the number M of normalization each sub-range mid point according to the sum M of all director interval midpoint _ifor r _i=M _i/ M, obtains directional spreding histogram with this, and trajectory direction histogram describes the statistics feature of trajectory direction, can be expressed as follows:

p ₃＝TDH＝(r ₁,r ₂,…,r _i,…,r _N-1,r _N)

Described center refers to the center position of every bar track after smoothing processing:

$p_4={\mathrm{\&Sigma;}}_{k=1}^N\stackrel{\&OverBar;}{t_k}/N$

Step 2.2: spectral clustering and multi-level distance metric

Adopt the spectral clustering implementation method based on Random Walk, data-oriented collection X=(x ₁, x ₂..., x _n), then:

W _ij＝exp(-dist(x _i,x _j)/2σ ²)

Wherein, dist (x _i, x _j) be distance metric, σ is standard variance, and the incidence matrix P of spectral clustering is transformed by matrix W and diagonal matrix D to draw, computing method are as follows:

P＝D ^-1/2WD ^-1/2

Wherein, diagonal matrix D

D＝diag(D ₁₁,…,D _ii,…,D _jj,…,D _NN)

In element D _iirepresent the summation of the i-th column element in similarity matrix and solve tag system eigenwert corresponding to it for P and proper vector just can complete spectral clustering;

For the needs of different levels spectral clustering, the present invention calculates the distance matrix dist (x in above formula according to different track characteristics _i, x _j), adopt Euclidean distance at ground floor and third layer, method is as follows:

E(i,j)＝||q _i-q _j|| ₂

Wherein, q _iand q _jrepresent the feature of i-th and jth bar track respectively, adopt at the second layer and calculate Pasteur's distance, method is as follows:

$H(i,j)={[1-\underset{b=1}{\overset{N}{\mathrm{\&Sigma;}}}\sqrt{{\mathrm{TDH}}_{\mathrm{ib}}{\mathrm{TDH}}_{\mathrm{jb}}}]}^{1/2}\&Element;\left[\mathrm{0,1}\right]$

Wherein TDH _iband TDH _jbexpression i-th article and b element in the corresponding TDH of jth article track respectively;

Step 3: intersection traffic properties of flow is analyzed and motion prediction

Step 3.1: sub-trajectory represents

The present invention adopts tapped delay line structure, vehicle movement track that is different for every bar length and that count different is expressed as the sub-trajectory of multiple regular length ξ, utilize tapped delay line structure, by copying ξ continuous sampling point, and the mode of each reach one creates sub-trajectory, ξ is a predefine parameter, and consider that between the stable description of initial trace and low operand trades off, this value is defined as follows:

$\mathrm{\&xi;}=\min (\sqrt{L},l_{\min })$

Wherein, L is the average length of track data collection, l _minit is the minimum length of track data collection;

Step 3.2: traffic stream characteristics analysis

Different vehicle sport mode shows that the vehicle of an approaching crossing is at the true stroke through crossing, comprises section, upstream, crossing that they arrive and specifically turns to, turning right, keep straight on, turn left;

The magnitude of traffic flow of the crossing point flow direction adopts the sum of track of vehicle in often kind of vehicle movement cluster to represent, because under all vehicle modes, temporal information is all consistent, therefore vehicle can be calculated by sub-trajectory representation through the average turning time of crossing, and the average turning time for different turn type can be expressed as the histogram of sub-trajectory arc length;

Step 3.3: motion prediction

Given componental movement track T _p, the present invention uses k nearest neighbor algorithm (k-NN) to be included to 12 track class set K _cinside the most general motor pattern, the similarity measurement in k-NN defines according to Euclidean distance, and method is as follows:

$d_c(T_p,K_c)=\sqrt{{\mathrm{\&Sigma;}}_{j\&Element;\mathrm{\&xi;}}{(T_p\left(j\right)-K_c\left(j\right))}^2}$

By Current vehicle driving trace T _pcompare to the corresponding track subclass of various motor pattern, calculate the movement locus probability P that vehicle is possible in the future _c:

$P_c=\frac{1/d_c}{{\mathrm{\&Sigma;}}_{c=1}^{12}1/d_c}$

Predict by selecting the type of gesture with maximum probability and complete the movement tendency that current kinetic vehicle may occur in the future the prediction of the current all vehicle movement trend in crossing simultaneously, realize the early warning of intersection vehicles sports safety.

Compared with prior art, the present invention has following clear superiority:

(1) the present invention proposes a kind of multi-level track Spectral Clustering for identifying different vehicle sport mode, is used for analyzing intersection traffic properties of flow and the early warning of vehicle movement safe prediction.

(2) root problem to be solved by this invention is when not having manual synchronizing, from original low-quality track of vehicle data, directly robustly find the motor pattern of vehicle inherence.

(3) the present invention extracts based on theory by local robust features, directly according to original low-quality track of vehicle data, analyses in depth the local geometric features near often in movement locus.Usually can produce a lot of error at track Data processing isolated point, use our bright institute using method can avoid the interference of these single point errors.

(4) the present invention can obtain the detailed traffic characterisitic parameter such as crossing point phase place (signalized crossing) magnitude of traffic flow and the journey time of each direction of motion vehicle through crossing, in this, as the important supplement of conventional traffic data.

(5) the present invention is by following the tracks of the travel path of all moving vehicles of current time, adopts next step behavior of the method prediction vehicle of traffic direction trajectory model coupling, the security risk that real-time early warning crossing may exist.

Accompanying drawing explanation

The general frame of Fig. 1 method involved in the present invention;

Fig. 2 crossing multi-level moving vehicle trajectory clustering result;

Fig. 3 a-3f intersection vehicles motion tracking result;

Fig. 4 a-4b initial trace preprocessing process;

Fig. 5 tapped delay line;

Fig. 6 a-6w is based on the vehicle sport mode recognition result of multi-level track spectral clustering framework;

The multi-level spectral clustering Performance comparision of Fig. 7;

Fig. 8 a-8f intersection traffic properties of flow is analyzed;

Fig. 9 intersection vehicles motion prediction.

Embodiment

Below in conjunction with drawings and Examples, the present invention will be further described.

The embodiment of the present invention realizes on the PC installing VC2008 and OpenCV2.4.5.

The process flow diagram of embodiment of the present invention method as shown in Figure 1, comprises the following steps:

Step 1: the intersection vehicles based on motion tracking move original rough track obtain, set up crossing extensive vehicle movement track data set.

Step 1.1: intersection vehicles motion initial trace gathers.

The true traffic video that the present invention is obtained by the video camera that the high building near crossing, Beijing sets up, tests the performance of multi-level spectral clustering framework, as shown in Figures 2 and 3.Adopt the object of which movement track algorithm that OpenCV realizes, process whole 17387 frame traffic video sequences, symbiosis becomes 1123 tracks, as shown in white portion in Fig. 3 d.In Fig. 3, each figure particular content is expressed as follows: a. sequence of video images; B. vehicle movement tracking results; C. background image; D. all movement locus be added in Background; E.3 trajectory diagram is tieed up; F.2 trajectory diagram is tieed up.

Step 1.2: original vehicle movement locus pre-service.

After pretreatment, still there are 997 tracks, as shown in fig. 4 a.In Fig. 4, each figure particular content is expressed as follows, and: a. is added to the crossing initial trace pre-processed results in Background; B. 2 dimension trajectory diagrams of crossing initial trace pre-processed results.

Step 2: the intersection vehicles motor pattern based on vehicle original rough track Multi-layer Spectral cluster learns.

Step 2.1: multi-level track characteristic extracts.

Step 2.2: indicated by traffic environment and signal control strategy, co-exists in 12 kinds of typical vehicle sport mode near this crossing.This and the present invention use the vehicle sport mode cluster result of multi-level spectral clustering framework identification to be consistent, as shown in Figure 6.In order to more effectively represent each track bunch, the present invention constructs template track more effectively can represent each track bunch.Template track is actual is exactly the minimum bunch heart of the Distance geometry of other all tracks in same cluster.

Given initial trace data as shown in fig. 4 a, the present invention represents the different tracks bunch of this intersection vehicles unique motion pattern by multi-level Spectral Clustering Multi-layer technology.Each track bunch is all represented by template track (the thick line section of the band arrow superposed in initial trace).Fig. 6 a, Fig. 6 b, Fig. 6 c are followed successively by the cluster result of different levels, have 4 respectively, 8,12 tracks bunch, wherein represent template track by different gray scale.4 tracks bunch in ground floor can specifically be expressed as Fig. 6 d-6g, and these 4 bunches of tracks have been further divided into 8 bunches in the second layer, as shown in Fig. 6 h-6o.4 bunches (Fig. 6 h-6k) are had again in third layer, to be become 8 bunches by Further Division, as shown in Fig. 6 p-6w in above 8 bunches.The node that in tree structure shown in final Fig. 2, existence 12 is different.In Fig. 6, each figure particular content is expressed as follows: a. ground floor cluster result (4 track bunch); B. second layer cluster result (8 track bunch); C. third layer cluster result (12 track bunch); D. ground floor turns 1; E. ground floor turns 2; F. ground floor keeps straight on 1; G. ground floor keeps straight on 2; H. the second layer is turned 1 direction 1; I. the second layer is turned 1 direction 2; J. the second layer is turned 2 directions 1; K. the second layer is turned 2 directions 2; L. the second layer is kept straight on 1 direction 1; M. the second layer is kept straight on 1 direction 2; N. the second layer is kept straight on 2 directions 1; O. the second layer is kept straight on 2 directions 2; P. third layer 1 direction 1 of turning is right; Q. third layer 1 direction 1 of turning is left; R. third layer 1 direction 2 of turning is left; S. third layer 1 direction 2 of turning is right; T. third layer 2 directions 1 of turning are left; U. third layer 2 directions 1 of turning are right; V. third layer 2 directions 2 of turning are right; W. third layer 2 directions 2 of turning are left.

In order to qualitative assessment typical vehicle motor pattern cluster result, the present invention use as follows closely and separation criteria (TSC) check the trajectory clustering effect of every one deck:

$\mathrm{TSC}=\frac{{\mathrm{\&Sigma;}}_{j=1}^k{\mathrm{\&Sigma;}}_{i=1}^n{R}_{\mathrm{ij}}*{\mathrm{dist}}^2(c_j,x_i)}{n*\min {\mathrm{dist}}^2(c_j,c_k)}$

Wherein c _jthe template track of jth bunch, in TSC measures bunch simultaneously tight ness rating and bunch between degree of separation.The less expression system performance of TSC numerical value is better.Fig. 7 (horizontal ordinate represents ever-increasing cluster level, and ordinate represents TSC value) clearly shows the increase along with the number of plies, and motor pattern is assembled better.

Step 3: intersection traffic properties of flow is analyzed and motion prediction.

Step 3.1: sub-trajectory represents.

Step 3.2: traffic stream characteristics analysis, the present invention mainly considers point garage's direction magnitude of traffic flow and an average travel time.

There are 12 kinds of typical vehicle motor patterns near selected crossing.According to the difference in arrived section, these typical modules can be classified as 4 kinds and arrive type, and often kind of an arrival type has 3 kinds of turn type further, comprises right-hand bend, craspedodrome, turns left, as shown in Figure 8 a-8d.Fig. 8 e and Fig. 8 f is expressed as the magnitude of traffic flow and the average travel time of given vehicle operating stroke.The magnitude of traffic flow that in Fig. 8 e, ordinate y illustrates a kind of pattern accounts for the number percent of total flow, and in Fig. 8 f, y illustrates the mean and variance of turn inside diameter time in every cluster vehicle operating stroke.In Fig. 8, each figure particular content is expressed as follows: 3 kinds of traffic directions of a. the 1st part of path; B. 3 kinds of traffic directions of the 2nd part of path; C. 3 kinds of traffic directions of the 3rd part of path; D. 3 kinds of traffic directions of the 4th part of path; E. the magnitude of traffic flow (horizontal ordinate represents 12 kinds of different vehicle sport mode cluster results) that respectively flows to of crossing; F. the average travel time (horizontal ordinate represents 12 kinds of different vehicle sport mode cluster results) that respectively flows to of crossing.Table 1 in detail sieve lists these numerical value.

The analysis of table 1 traffic stream characteristics

Step 3.3: intersection vehicles motion prediction.

Fig. 9 illustrates concrete motion prediction example, the current kinetic track of the wherein detected vehicle of thick black solid line representative (adopting grey oval marks), thin black represented by dotted arrows prediction locus.The percentage whiteness number on thin black dotted line side represents the probability of detected vehicle possibility motor pattern in crossing.Vehicle enters this traffic scene from section, right side in fig .9, only remains the prediction locus corresponding to first 3 of probable value.When vehicle turns left, a kind of probable value of prediction locus is only had to be increased to 99.42%.Similar, the prediction of the current all vehicle movement trend in crossing can be completed, and then realize the early warning of intersection vehicles sports safety.

Last it is noted that above example only in order to illustrate the present invention and and unrestricted technical scheme described in the invention; Therefore, although this instructions with reference to above-mentioned example to present invention has been detailed description, those of ordinary skill in the art should be appreciated that and still can modify to the present invention or equivalent to replace; And all do not depart from technical scheme and the improvement thereof of the spirit and scope of invention, it all should be encompassed in the middle of right of the present invention.

Claims (1)

1. the intersection traffic properties of flow based on track data is analyzed and a vehicle movement Forecasting Methodology, it is characterized in that, comprises the steps:

Step 1: the intersection vehicles based on motion tracking move original rough track obtain, set up crossing extensive vehicle movement track data set;

Step 1.1: intersection vehicles motion initial trace gathers

Adopt the tracking based on image block in OpenCV, automatically extract crossing moving vehicle initial trace data, and be expressed as vehicle movement point sequence T:

T＝{t ₁，t ₂，…，t _i，…，t _n}

＝{(x ₁，y ₁)，(x ₂，y ₂)，…，(x _i，y _i)，…，(x _n，y _n)}

And track step sequence S:

S＝{s ₁，s ₂，…，s _i，…，s _n-1}

＝{(δx ₁，δy ₁)，…，(δx _i，δy _i)，…，(δx _n-1，δy _n-1)}

Wherein, t _i=(x _i, y _i) represent the position at moving vehicle i-th sampled point center, s _i=(δ x _i, δ y _i)=(x _i+1-x _i, y _i+1-y _i) represent the deviation of neighbouring sample dot center, n represent track of vehicle comprise the sum of sampled point;

Step 1.2: original vehicle movement locus pre-service

Carry out following smoothing processing for every bar track: (1) considers sample noise, as the distance between fruit dot is enough little, just continuous print point combined and replaced by first point; (2) complete deletion length is less than the short track of predefine threshold value; (3) the smoothing process of mean filter is adopted, to retain prototype structure essential characteristic (selecting level and smooth step number w=7 in experiment):

$\stackrel{\&OverBar;}{t_i}=\frac{1}{w}\underset{j=i-w/2}{\overset{i+w/2}{\mathrm{\&Sigma;}}}{t}_j$

Step 2: the intersection vehicles motor pattern based on vehicle original rough track Multi-layer Spectral cluster learns

Step 2.1: multi-level track characteristic extracts

The present invention uses linearity and flexibility, trajectory direction histogram and three kinds, center feature to carry out the local feature of Integrative expression track respectively;

Described linearity and flexibility refer to the tolerance that vehicle operating orientation average changes, if α _irepresent s _iwith s _i+1orientation angle change between step, establishes simultaneously and is changed to positive dirction left;

Linearity is defined as follows:

$p_1=\frac{1}{l_w-1}\underset{i=j}{\overset{j+l_w-2}{\mathrm{\&Sigma;}}}\cos {\mathrm{\&alpha;}}_j$

Flexibility is defined as follows:

$p_2=\frac{1}{l_w-1}\underset{i=j}{\overset{j+l_w-2}{\mathrm{\&Sigma;}}}\sin {\mathrm{\&alpha;}}_j$

Described trajectory direction histogram TDH is a kind of feature representation method of novel description trajectory direction.First, the deflection β of a jth point in track is calculated by the following method _j:

${\mathrm{\&beta;}}_j=\left\{\begin{array}{c}\arctan \left(\frac{\mathrm{dy}}{\mathrm{dx}}\right),\mathrm{dx}>0\\ \arctan \left(\frac{\mathrm{dy}}{\mathrm{dx}}\right)+\mathrm{\&pi;},\mathrm{dx}<0,\mathrm{dy}>0\\ \arctan \left(\frac{\mathrm{dy}}{\mathrm{dx}}\right)-\mathrm{\&pi;},\mathrm{dx}<0,\mathrm{dy}<0\end{array}\right.$

Wherein, (dx is required _j) ²+ (dy _j) ²≠ 0 and β _j∈ [-π, + π), then, by interval [-π, + π) be evenly divided into N number of equal-sized director interval, and by all points in identical strip path curve according to the different mappings of deflection in corresponding director interval, finally, carry out the number M of normalization each sub-range mid point according to the sum M of all director interval midpoint _ifor r _i=M _i/ M, obtains directional spreding histogram with this, and trajectory direction histogram describes the statistics feature of trajectory direction, can be expressed as follows:

p ₃＝TDH＝(r ₁，r ₂，…，r _i，…，r _N-1，r _N)

Described center refers to the center position of every bar track after smoothing processing:

$p_4={\mathrm{\&Sigma;}}_{k=1}^N\stackrel{\&OverBar;}{t_k}/N$

Step 2.2: spectral clustering and multi-level distance metric

Adopt the spectral clustering implementation method based on Random Walk, data-oriented collection X=(x ₁, x ₂..., x _n), then:

W _ij＝exp(-dist(x _i，x _j)/2σ ²)

Wherein, dist (x _i, x _j) be distance metric, σ is standard variance, and the incidence matrix P of spectral clustering is transformed by matrix W and diagonal matrix D to draw, computing method are as follows:

P＝D ^-1/2WD ^-1/2

Wherein, diagonal matrix D

D＝diag(D ₁₁，…，D _ii，…，D _jj，…，D _NN)

In element D _iirepresent the summation of the i-th column element in similarity matrix and solve tag system eigenwert corresponding to it for P and proper vector just can complete spectral clustering;

For the needs of different levels spectral clustering, the present invention calculates the distance matrix dist (x in above formula according to different track characteristics _i, x _j), adopt Euclidean distance at ground floor and third layer, method is as follows:

E(i，j)＝||q _i-q _j|| ₂

Wherein, q _iand q _jrepresent the feature of i-th and jth bar track respectively, adopt at the second layer and calculate Pasteur's distance, method is as follows:

$H(i,j)={[1-\underset{b=1}{\overset{N}{\mathrm{\&Sigma;}}}\sqrt{{\mathrm{TDH}}_{\mathrm{ib}}{\mathrm{TDH}}_{\mathrm{jb}}}]}^{1/2}\&Element;\left[\mathrm{0,1}\right]$

Wherein TDH _iband TDH _jbexpression i-th article and b element in the corresponding TDH of jth article track respectively;

Step 3: intersection traffic properties of flow is analyzed and motion prediction

Step 3.1: sub-trajectory represents

The present invention adopts tapped delay line structure, vehicle movement track that is different for every bar length and that count different is expressed as the sub-trajectory of multiple regular length ξ, utilize tapped delay line structure, by copying ξ continuous sampling point, and the mode of each reach one creates sub-trajectory, ξ is a predefine parameter, and consider that between the stable description of initial trace and low operand trades off, this value is defined as follows:

$\mathrm{\&xi;}=\min (\sqrt{L},l_{\min })$

Wherein, L is the average length of track data collection, l _minit is the minimum length of track data collection;

Step 3.2: traffic stream characteristics analysis

Different vehicle sport mode shows that the vehicle of an approaching crossing is at the true stroke through crossing, comprises section, upstream, crossing that they arrive and specifically turns to, turning right, keep straight on, turn left;

The magnitude of traffic flow of the crossing point flow direction adopts the sum of track of vehicle in often kind of vehicle movement cluster to represent, because under all vehicle modes, temporal information is all consistent, therefore vehicle can be calculated by sub-trajectory representation through the average turning time of crossing, and the average turning time for different turn type can be expressed as the histogram of sub-trajectory arc length;

Step 3.3: motion prediction

Given componental movement track T _p, the present invention uses k nearest neighbor algorithm (k-NN) to be included to 12 track class set K _cinside the most general motor pattern, the similarity measurement in k-NN defines according to Euclidean distance, and method is as follows:

$d_c(T_p,K_c)=\sqrt{{\mathrm{\&Sigma;}}_{j\&Element;\mathrm{\&xi;}}{(T_p\left(j\right)-K_c\left(j\right))}^2}$

By Current vehicle driving trace T _pcompare to the corresponding track subclass of various motor pattern, calculate the movement locus probability P that vehicle is possible in the future _c:

$P_c=\frac{1/d_c}{{\mathrm{\&Sigma;}}_{c=1}^{12}1/d_c}$

Predict by selecting the type of gesture with maximum probability and complete the movement tendency that current kinetic vehicle may occur in the future the prediction of the current all vehicle movement trend in crossing simultaneously, realize the early warning of intersection vehicles sports safety.