CN104657424A - Clustering method for interest point tracks under multiple temporal and spatial characteristic fusion - Google Patents

Abstract

The invention discloses a clustering method for interest point tracks under multiple temporal and spatial characteristic fusion. The clustering method comprises the following steps: firstly, acquiring interest points where users stay according to the temporal and spatial span between continuous coordinate points of the tracks, and segmenting the tracks into a plurality of track segments through the interest points; secondly, calculating the difference of the tracks in the space, the time, the speed and the direction to comprehensively judge the similarity between two tracks; finally, performing density clustering on the tracks based on an OPTICS method, and cutting off clustering clusters with sparse track quantity. According to the clustering method disclosed by the invention, the tracks are converted into an interest point sequence; the tracks are clustered by comprehensively considering the speed, the direction and temporal and spatial characteristics of the track segment between every two continuous interest points, so a significant cluster is selected and further the track clustering form which can comprehensively reflect global importance is obtained. Provided by experimental results, the method retains original temporal and spatial characteristics and the moving attribute of the tracks, can comprehensively reflect the motion and behavior mode of a mobile object and is high in clustering accuracy.

Description

Point of interest method of trajectory clustering under a kind of multi-space Fusion Features

Technical field

The invention belongs to trajectory clustering algorithm field, particularly relate to the point of interest method of trajectory clustering under a kind of multi-space Fusion Features.

Background technology

Along with the development of satellite, internet and tracking equipment, the track data of a large amount of mobile object is captured, as vehicle move, animal is moved, typhoon moves towards, personnel move.These track datas accumulated in a large number have recorded the position of mobile object and the records series of time, have contained abundant space-time knowledge, have had huge using value.By analyzing track data, contribute to studying every field such as human behavioral mode, communication and logistics, emergency evacuation management, Animal behaviour, the marketing, computational geometry and analog simulations.By carrying out cluster analysis to various space-time trajectory data, the similarity in space-time trajectory data and off-note can be extracted, and contribute to finding wherein significant spatiotemporal mode.

In recent years, the extensive concern of Chinese scholars is caused based on the behavior analysis method of track.But these methods do not carry out cluster in conjunction with the multiple moving characteristic information such as speed, direction mostly.In addition, track is split as orbit segment by existing method, carries out cluster to orbit segment, and cluster result cannot from the space-time characteristic of overall angle reflection track and movement tendency.

Summary of the invention

The object of the embodiment of the present invention is to provide the point of interest method of trajectory clustering under a kind of multi-space Fusion Features, be intended to solve overall space-time similarity and the local space time's similarity that current trajectory clustering algorithm does not consider track, cause urban activity trajectory clustering to be difficult to embody the problem of social activities similarity.

The present invention is achieved in that the point of interest method of trajectory clustering under a kind of multi-space Fusion Features comprises:

Step one, the point of interest stopped according to the Time and place span acquisition user between track continuous coordinate points, be divided into many orbit segments by point of interest by track;

Step 2, by calculating the similarity that the difference of orbit segment on space, time, speed and direction is come between comprehensive descision two tracks;

Step 3, based on OPTICS method, Density Clustering is carried out to track, and cut out the sparse clustering cluster of tracking quantity.

Further, the concrete grammar calculating orbit segment space-time similarity is:

Step one, based on space quaternary tree index, with orbit segment each in track S for query object, to the immediate corresponding orbit segment of S each orbit segment space length in quick obtaining track R, if the space length between two orbit segments is less than threshold value, then they form orbit segment to SP;

Step 2, orbit segment in same SP carry out follow-up Similarity measures, and the Similarity measures between orbit segment comprises 4 aspects: spatial simlanty spatialSIM, chronotaxis tempoSIM, directional similarity OrientSIM and speed similarity velocitySIM.

Further, in described calculating orbit segment space-time similarity, to be the difference value of two orbit segment U, V point of interest at whole story centers be spatialSIM:

Δd _u，v＝|p _u，s-p _v，s|+|p _u，e-p _v，e|

Chronotaxis tempoSIM comprises the time difference of two orbit segment coordinate points at the whole story and orbit segment difference interval time is:

Δt _u，v＝|t _u，s-t _v，s|+|t _u，e-t _v，e|+|Δt _u-Δt _v|

；

+|Δtp _u，s-Δtp _v，s|+|Δtp _u，e-Δtp _v，e|

Directional similarity OrientSIM is:

Δθ _u，v＝|θ _u-θ _v|；

Speed similarity velocitySIM is:

$\mathrm{\Δ}{\stackrel{\‾}{V}}_{u,v}=|{\stackrel{\‾}{V}}_u-{\stackrel{\‾}{V}}_v|;$

The inverse of LCSS is adopted to obtain similarity (the Δ L of each sampled point between orbit segment _{u, v});

Time-space matrix formula between track U and V is:

$\begin{array}{c}\mathrm{DIS}(U,V)=\frac{\exp \left\{\frac{I\left[U\right]+I\left[V\right]}{2}\right\}}{\min \left(\right|U|,|V\left|\right)}\&times;\\ {\mathrm{\&Sigma;}}_{<u,v>,u\&Element;U,v\&Element;V}\frac{\mathrm{\&Delta;}{d}_{u,v}\mathrm{\&Delta;}{L}_{u,v}\mathrm{\&Delta;}{\mathrm{\&theta;}}_{u,v}{\mathrm{\&Delta;t}}_{u,v}}{\min \left(\right|u|,|v\left|\right)}\end{array}$

Further, the time-space matrix formula between track U and V is utilized, based on OPTICS algorithm to trajectory clustering, the clustering cluster that tracking quantity is less must be filtered out, reduce a large amount of broken cluster, allow final trajectory clustering result can reflect the trunk path profile with important global sense, concrete grammar is:

If the track number comprised in clustering cluster C is bunch radix n _cb, given threshold tau, definition bunch radix n _bbe called remarkable bunch higher than τ clustering cluster, remaining cluster result is non-significant bunch, finally cropped fall.

Track is converted to point of interest sequence by the present invention, considers the speed of the orbit segment formed between continuous point of interest, direction and space-time characterisation, carries out trajectory clustering, and picks out remarkable bunch to obtain the trajectory clustering form that can reflect overall importance.Experimental result shows, the method remains the original space-time of track and mobile attribute characteristic in cluster, and more fully can reflect motion and the behavior pattern of mobile object, cluster accuracy is high.

Accompanying drawing explanation

Fig. 1 is the point of interest method of trajectory clustering process flow diagram under the multi-space Fusion Features that provides of the embodiment of the present invention;

Fig. 2 is remarkable bunch after the cluster of the MSFF that the embodiment of the present invention provides;

Fig. 3 is the LUCC clustering result figure that the embodiment of the present invention provides;

Fig. 4 is remarkable bunch after the LUCC clustering that provides of the embodiment of the present invention;

Fig. 5 is the cluster accuracy rate comparison diagram that the embodiment of the present invention provides.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with embodiment, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Below in conjunction with drawings and the specific embodiments, application principle of the present invention is further described.

Fig. 1 shows the point of interest method of trajectory clustering flow process under multi-space Fusion Features of the present invention, and as shown in the figure, the present invention is achieved in that the point of interest method of trajectory clustering under a kind of multi-space Fusion Features comprises:

S101, the point of interest stopped according to the Time and place span acquisition user between track continuous coordinate points, be divided into many orbit segments by point of interest by track;

S102, by calculating the similarity that the difference of orbit segment on space, time, speed and direction is come between comprehensive descision two tracks;

S103, based on OPTICS method, Density Clustering is carried out to track, and cut out the sparse clustering cluster of tracking quantity.

Further, the concrete grammar calculating orbit segment space-time similarity is:

Step one, based on space quaternary tree index, with orbit segment each in track S for query object, to the immediate corresponding orbit segment of S each orbit segment space length in quick obtaining track R, if the space length between two orbit segments is less than threshold value, then they form orbit segment to SP;

Step 2, orbit segment in same SP carry out follow-up Similarity measures, and the Similarity measures between orbit segment comprises 4 aspects: spatial simlanty spatialSIM, chronotaxis tempoSIM, directional similarity OrientSIM and speed similarity velocitySIM.

Further, in described calculating orbit segment space-time similarity, to be the difference value of two orbit segment U, V point of interest at whole story centers be spatialSIM:

Δd _u，v＝|p _u，s-p _v，s|+|p _u，e-p _v，e|

Chronotaxis tempoSIM comprises the time difference of two orbit segment coordinate points at the whole story and orbit segment difference interval time is:

Δt _u，v＝|t _u，s-t _v，s|+|t _u，e-t _v，e|+|Δt _u-Δt _v|

；

+|Δtp _u，s-Δtp _v，s|+|Δtp _u，e-Δtp _v，e|

Directional similarity OrientSIM is:

Δθ _u，v＝|θ _u-θ _v|；

Speed similarity velocitySIM is:

$\mathrm{\&Delta;}{\stackrel{\&OverBar;}{V}}_{u,v}=|{\stackrel{\&OverBar;}{V}}_u-{\stackrel{\&OverBar;}{V}}_v|;$

The inverse of LCSS is adopted to obtain similarity (the Δ L of each sampled point between orbit segment _{u, v});

Time-space matrix formula between track U and V is:

$\begin{array}{c}\mathrm{DIS}(U,V)=\frac{\exp \left\{\frac{I\left[U\right]+I\left[V\right]}{2}\right\}}{\min \left(\right|U|,|V\left|\right)}\&times;\\ {\mathrm{\&Sigma;}}_{<u,v>,u\&Element;U,v\&Element;V}\frac{\mathrm{\&Delta;}{d}_{u,v}\mathrm{\&Delta;}{L}_{u,v}\mathrm{\&Delta;}{\mathrm{\&theta;}}_{u,v}{\mathrm{\&Delta;t}}_{u,v}}{\min \left(\right|u|,|v\left|\right)}\end{array}$

Further, the time-space matrix formula between track U and V is utilized, based on OPTICS algorithm to trajectory clustering, the clustering cluster that tracking quantity is less must be filtered out, reduce a large amount of broken cluster, allow final trajectory clustering result can reflect the trunk path profile with important global sense, concrete grammar is:

If the track number comprised in clustering cluster C is bunch radix n _cb, given threshold tau, definition bunch radix n _bbe called remarkable bunch higher than τ clustering cluster, remaining cluster result is non-significant bunch, finally cropped fall.

1, concept and definition

Track: in plane space pass in time and the point sequence that adopts continuously is called track.Track TR _idefinition: TR _i={ p ₁, p ₂..., p _k, wherein p _k={ x _k, y _k, t _k, they represent two-dimensional space coordinate and the employing time of this point respectively.

Point of interest: suppose existence one continuous sampling point set, this point concentrates the distance of any point-to-point transmission to be no more than threshold epsilon _p, and sampling number s is greater than threshold epsilon _t, the minimum enclosed rectangle that in this set, all sampled points build, is called point of interest, and it represents the behaviors such as user gets on the bus in this position, get off or waits traffic lights, and implies user behavior semanteme or traffic semantic information.Therefore, the space-time similarity measurement between track can be similar to the adjacency be converted between two track points of interest: spatially more close between the point of interest forming two tracks, space overlap degree is higher, and overlapping point of interest is more, then they are more similar.But except considering the spacial proximity between point of interest, between track, Similarity measures also will consider the factors such as the similarity between direction, speed and other sampled points.

Orbit segment: the sub-trajectory be made up of continuous two points of interest and the sampled point between them, is expressed as: SubTrajectorys={p ₁..., p _k(1≤s≤k), the total number of sample points of k track belonging to this orbit segment.

Orbit segment speed: the speed of orbit segment u is that minimum speed, maximal rate and the average velocity in all sampled points of this orbit segment is weighed:

$\begin{array}{c}{\stackrel{\&OverBar;}{V}}_u=(1-{\mathrm{\&omega;}}_m-{\mathrm{\&omega;}}_a)v_{\min }+{\mathrm{\&omega;}}_m{v}_{\max }\\ +{\mathrm{\&omega;}}_a\frac{v_i+v_{i+1}+...+v_j}{j-i+1}\end{array}---\left(1\right)$

Wherein, ω _m+ ω _a≤ 1, v _minfor orbit segment medium velocity minimum, v _maxfor orbit segment medium velocity mxm., i and j is respectively the subscript of this orbit segment sampled point.

Orbit segment direction: be the angle formed between orbit segment u point of interest at the whole story, also claims direction of motion angle:

${\mathrm{\&theta;}}_u=\arctan \left[\frac{(y_{u,e}-y_{u,s})}{(x_{u,e}-x_{u,s})}\right]---\left(2\right)$

Wherein, (x _{u, s}, y _{u, s}) orbit segment starting point, (x _{u, e}, y _{u, e}) be orbit segment terminal.

The track point of interest residence time: for user distinguishes the resident time at orbit segment point of interest at the whole story, indirectly reflect the busy extent of mobile object, Δ tp can be expressed as _{u, s}with Δ tp _{u, e}.

Orbit segment interval time: be the time interval between the whole story in rail stage point of interest, have expressed the traffic of user between these two points of interest and to travel frequently situation and the mode of transportation that mobile object adopts.It can be expressed as:

Δt _u＝t _u，s-t _u，e(3)

Wherein, t _{u, s}and t _{u, e}represent that orbit segment u samples the corresponding time at whole story respectively.

2, track space-time cluster

2.1 orbit segment similarity measurements

Longest common subsequence (Longest common subse quence, LCSS): refer in two or more sequences the longest common subsequence existed.For space-time track, calculate its longest common subsequence and be converted into LCSS distance and can weigh similarity degree between track.LCSS value is obtained by recursive fashion:

$\mathrm{LCSS}(R,S)=\left\{\begin{array}{c}0,\mathrm{if}(m=n=0)\\ \mathrm{LCSS}(\mathrm{Rest}\left(R\right),\mathrm{Rest}\left(S\right))+1,\mathrm{if}|r_{1,x}-s_{1,x}|\&le;\mathrm{\&delta;}\&cap;|r_{1,y}-s_{1,y}|\&le;\mathrm{\&psi;};\\ \max \left\{\mathrm{LCSS}(\mathrm{Rest}\left(R\right),S)\mathrm{LCSS}(R,\mathrm{Rest}\left(S\right))\right\},\mathrm{otherwise}.\end{array}\right.---\left(3\right)$

In formula, LCSS (R, S) represents the similarity between track R and S.Suppose that track compares at two-dimensional space, δ and ψ represents the similar threshold value in x-axis and y-axis respectively, and when Diff E is less than δ and Diff N is less than ψ, think that this is similar to sampled point, LCSS value adds 1, and other every meanings as hereinbefore.When track record count be all 0 time, LCSS (R, S) is 0.The present embodiment is quoted LCSS and is calculated orbit segment space-time similarity.

First, based on space quaternary tree index, with orbit segment each in track S for query object, with S each orbit segment space length (adopting in orbit segment space length in the heart to represent) immediate corresponding orbit segment in quick obtaining track R.If the space length between two orbit segments is less than threshold value, then they form orbit segment to (Sub-Trajectory Pair, SP).This method avoid in LCSS and search point of interest nearest between two tracks by the method for recurrence, reduce the time complexity of Similarity measures.Secondly, the orbit segment in same SP carries out follow-up Similarity measures.Similarity measures between orbit segment comprises 4 aspects: spatial simlanty spatialSIM, chronotaxis tempoSIM, directional similarity OrientSIM and speed similarity velocitySIM.Wherein, spatialSIM is the difference value of two orbit segment point of interest at whole story centers, that is:

Δd _u，v＝|p _u，s-p _v，s|+|p _u，e-p _v，e| (5)

Wherein, Δ d _{u, v}for the space length between orbit segment u and v, p _{u, s}for the center of the initial point of interest of orbit segment u, p _{u, e}for the center of the terminal point of interest of orbit segment v.In addition, similarity (the Δ L of each sampled point between orbit segment will also be further considered _{u, v}), adopt the inverse of LCSS to obtain this value here.

Chronotaxis tempoSIM comprises time difference and orbit segment difference interval time of two orbit segment coordinate points at the whole story, that is:

Δt _u，v＝|t _u，s-t _v，s|+|t _u，e-t _v，e|+|Δt _u-Δt _v|

(6)

+|Δtp _u，s-Δtp _v，s|+|Δtp _u，e-Δtp _v，e|

Wherein, t _{u, s}for the starting point time of orbit segment u, t _{u, e}for the terminal time of orbit segment u.

Directional similarity OrientSIM is:

Δθ _u，v＝|θ _u-θ _v| (1)

Speed similarity velocitySIM is:

$\mathrm{\&Delta;}{\stackrel{\&OverBar;}{V}}_{u,v}=|{\stackrel{\&OverBar;}{V}}_u-{\stackrel{\&OverBar;}{V}}_v|---\left(2\right)$

Finally, above formula is merged, the time-space matrix formula between track U and V can be obtained:

$\begin{array}{c}\mathrm{DIS}(U,V)=\frac{\exp \left\{\frac{I\left[U\right]+I\left[V\right]}{2}\right\}}{\min \left(\right|U|,|V\left|\right)}\&times;\\ {\mathrm{\&Sigma;}}_{<u,v>,u\&Element;U,v\&Element;V}\frac{\mathrm{\&Delta;}{d}_{u,v}\mathrm{\&Delta;}{L}_{u,v}\mathrm{\&Delta;}{\mathrm{\&theta;}}_{u,v}{\mathrm{\&Delta;t}}_{u,v}}{\min \left(\right|u|,|v\left|\right)}\end{array}---\left(3\right)$

Wherein, I [U] represents the orbit segment quantity not belonging to any SP at U, and <u, v> form a SP, and they are respectively the orbit segment in track U and V.

2.2 based on the trajectory clustering of OPTICS

Utilize track similarity formula (9), based on OPTICS (Ordering Points To Identify theClustering Structure) algorithm to trajectory clustering.

Definition ξ neighborhood N _ξ(R): for track R, given proximity threshold ξ, if there is track S, meets N _ξ(R)={ S ∈ D|DIS (R, S)≤λ, R ≠ S}.Wherein, D is all track data set.Track neighborhood, in order in OPTICS cluster, judges the current density of every bar track, and then gathers track larger for density for same group.

The clustering cluster that those tracking quantity are less must be filtered out, reduce a large amount of broken cluster, allow final trajectory clustering result can reflect the trunk path profile with important global sense.

If the track number comprised in clustering cluster C is bunch radix n _cb, given threshold tau, carry out as given a definition:

Remarkable bunch: C _pro={ C|C ∈ O ∩ n _cb> τ }, wherein, O is the result set of cluster.I.e. bunch radix n _bremarkable bunch is called higher than τ clustering cluster.Remaining cluster result is non-significant bunch, finally cropped fall.

The method better can be filled into most of unessential cluster, global space divide different on highlight the coverage of important cluster, and other similar track Density Clustering methods cannot accomplish this point.

3, clustering schemes is implemented

Server hardware in this programme deployed environment is CPU (CORE 2 DUO 2.8GH), internal memory 8GB, adopts Visual Studio 2010 to programme, and adopts MySQL to store and retrieval track data.Adopt the taxi data set in wuchang, wuhan district in February, 2010 to April as experimental data, totally 10835 tracks, the sampled point of every bar track includes latitude and longitude coordinates, sampling time.Clustering method is unified adopts OPTICS.

3.1 Cluster Validities compare

Method in this paper is referred to as MSFF, and comparative approach is adopt LUCC for the track similarity calculation method of initial trace spatial coordinated information, referred to as LUCC.Fig. 2-Fig. 4 respectively illustrates the trajectory clustering spatial distribution map of MSFF method and LUCC method, in figure, every bar line represents 100 and does not distinguish speed, direction and the track of time, be less than 100 also represent with a line segment, and be with coloured lines to be cluster, grey lines are non-significant bunch result.

Fig. 2 is remarkable bunch of spatial distribution result after this method MSFF cluster.In this figure, track data is divided into 76 clustering cluster.After track is divided into each point of interest by the present embodiment, the track that speed but point of interest location similar with direction differs greatly on same turnpike road originally, can not be aggregated to same clustering cluster.Lower owing to being positioned at the speed of mobile object and the similarity in direction on limb road, be mostly in track outside major trunk roads by as non-significant bunch process.

The cluster result of LUCC is shown as in Fig. 3.OPTICS is through arameter optimization, and the track after its process is divided into 282 classes.In general, although can embody the main distribution of urban traffic, because number of clusters is numerous, many collateral branch path and primary highway have gathered one piece, are difficult to the distribution embodying city main traffic artery.

Fig. 4 is space distribution LUCC method Fig. 3 cluster result basis filtering out non-significant bunch.At this moment OPTICS cluster result has filtered out the very few clustering cluster of a large amount of tracking quantity, and final track is only divided into 12 classes, only can embody the distribution of urban parts main traffic artery.

Visible, the space distribution of this method gained trajectory clustering result reflects the most main flow condition of wagon flow in city, and good behaviour has gone out trunk thoroughfare, city, and Clustering Effect is remarkable.

3.2 cluster accuracys rate compare

Carry out MSFF and LUCC Measures compare:

$\mathrm{\&eta;}=1-\frac{{\mathrm{dist}}_m\left(i\right)}{{\mathrm{\&Sigma;}}_{n=1,n\&NotEqual;m}^K{\mathrm{dist}}_m\left(i\right)},$ Wherein, ${\mathrm{dist}}_n\left(i\right)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;C_n}\mathrm{DIS}(i,j)}{\left|C_n\right|}---\left(4\right)$

Dist _ni () is the distance between track i and clustering cluster n.η is larger, and cluster accuracy rate is higher.

Scheme implementation result as shown in Figure 5.Along with the increase of proximity threshold ξ in OPTICS method, the cluster accuracy rate of MSFF is all obviously better than LUCC method.This is because MSFF has considered the room and time various features between point of interest and sampled point, also taken into account social semantic information, therefore in each cluster, the space time information consistance of track is high, can express the similar track that space-time granularity is more careful; And LUCC method only considers the spatial simlanty under global context, the consistance of the space time information therefore in each cluster is relatively low, and the track in same cluster can not have the mobile object of social mobile Semantic Similarity by correction, and cluster applicable surface is narrow.

4, conclusion

Track space-time cluster migrates the fields such as pattern, magnitude of traffic flow differentiation, geographical social recommendation animal have significant application value.Track is converted to point of interest sequence herein, considers the speed of the orbit segment formed between continuous point of interest, direction and space-time characterisation, carry out trajectory clustering, and pick out remarkable bunch to obtain the trajectory clustering form that can reflect overall importance.Experimental result shows, the method remains the original space-time of track and mobile attribute characteristic in cluster, and more fully can reflect motion and the behavior pattern of mobile object, cluster accuracy is high.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (4)

1. the point of interest method of trajectory clustering under a multi-space Fusion Features, it is characterized in that, point of interest method of trajectory clustering under this multi-space Fusion Features utilizes movable information and the point of interest location information of track, more accurately expresses the track set with social action similarity;

Specifically comprise the following steps:

Step one, the point of interest stopped according to the Time and place span acquisition user between track continuous coordinate points, be divided into many orbit segments by point of interest by track;

Step 2, by calculating the similarity that the difference of orbit segment on space, time, speed and direction is come between comprehensive descision two tracks;

Step 3, based on OPTICS method, Density Clustering is carried out to track, and cut out the sparse clustering cluster of tracking quantity.

2. the point of interest method of trajectory clustering under multi-space Fusion Features as claimed in claim 1, it is characterized in that, make full use of movable information and the point of interest location information of track, express the track set with social action similarity, the concrete grammar calculating orbit segment space-time similarity is:

Step one, based on space quaternary tree index, with orbit segment each in track S for query object, obtain in track R to the immediate corresponding orbit segment of S each orbit segment space length, the space length between two orbit segments is less than threshold value, then form orbit segment to SP;

Step 2, orbit segment in same SP carry out follow-up Similarity measures, and the Similarity measures between orbit segment comprises 4 aspects: space with property spatialSIM, time with property tempoSIM, direction with property OrientSIM and speed similarity velocitySIM.

3. the point of interest method of trajectory clustering under multi-space Fusion Features as claimed in claim 2, it is characterized in that, make full use of movable information and the point of interest location information of track, express the track set with social action similarity, calculate in orbit segment space-time similarity, to be the difference value of two orbit segment U, V point of interest at whole story centers be spatialSIM:

Δd _u，v＝|p _u，s-p _v，s|+|p _u，e-p _v，e|

Chronotaxis tempoSIM comprises the time difference of two orbit segment coordinate points at the whole story and orbit segment difference interval time is:

Δt _u，v＝|t _u，s-t _v，s|+|t _u，e-t _v，e|+|Δt _u-Δt _v|

；

+|Δtp _u，s-Δtp _v，s|+|Δtp _u，e-Δtp _v，e|

Directional similarity OrientSIM is:

Δθ _u，v＝|θ _u-θ _v|；

Speed similarity velocitySIM is:

$\mathrm{\&Delta;}{\stackrel{\&OverBar;}{V}}_{u,v}=|{\stackrel{\&OverBar;}{V}}_u-{\stackrel{\&OverBar;}{V}}_v|;$

The inverse of LCSS is adopted to obtain similarity (the Δ L of each sampled point between orbit segment _{u, v});

Time-space matrix formula between track U and V is:

$\begin{array}{c}\mathrm{DIS}(U,V)=\frac{\exp \left\{\frac{I\left[U\right]+I\left[V\right]}{2}\right\}}{\min \left(\right|U|,|V\left|\right)}\&times;\\ {\mathrm{\&Sigma;}}_{\&lang;u,v\&rang;,u\&Element;U,v\&Element;V\frac{\mathrm{\&Delta;}{d}_{u,v}\mathrm{\&Delta;}{L}_{u,v}\mathrm{\&Delta;}{\mathrm{\&theta;}}_{u,v}\mathrm{\&Delta;}{t}_{u,v}}{\min \left(\right|u|,|v\left|\right)}}\end{array}.$

4. the point of interest method of trajectory clustering under multi-space Fusion Features as claimed in claim 1, it is characterized in that, utilize the time-space matrix formula between track U and V, based on OPTICS algorithm to trajectory clustering, allow final trajectory clustering result can reflect the trunk path profile with important global sense, concrete grammar is: the track number comprised in clustering cluster C is bunch radix n _cb, given threshold tau, definition bunch radix n _bbe called remarkable bunch higher than τ clustering cluster, remaining cluster result is non-significant bunch, finally cropped fall.