2. Motivations Analyzing objects movement to identify frequently moving paths at different scale can improve a large set of services such as: Traffic management. Location-based service. Customer behavior analysis. CCTV Surveillance. 2 15 March 2010
4. How to realize Using Trajectory segment as the main interest unit Clustering all the Trajectory segment based on density control Extract the cluster multi-granularity 4 15 March 2010
5. Trajectory definition Definition 1.A trajectory is a sequence of multi-dimensional points. Definition 2.A cluster is a set of similar sub-trajectories. 5 15 March 2010
6. Method 6 Select the set of interesting trajectories Corner detection Segment the trajectories Get the augmented order of the sub-trajectories Frechet distance measure Extract the cluster with different granularity Visualization 15 March 2010
7. Trajectory segmentation Scans the point sequence of a trajectory and selects candidate corner points according to the calculation of the open angle Remove redundant candidates. 7 15 March 2010
11. Distance measure for sub-trajectory clustering Fréchet Distance Discrete Fréchet Distance 9 15 March 2010
12. Density-based clustering A cluster: a maximal set of density-connected points Discover clusters of arbitrary shape in spatial databases with noise OPTICS: ordering points to identify the clustering structure It computes an augmented clustering-ordering for automatic cluster analysis. Based on this idea, two values need to be introduced: Core distance Reachability distance 10 15 March 2010
13. Density-based clustering OPTICS Algorithm create an ordering of a database, additionally store the core-distance and a suitable reachability-distance for each points. Then such information is used to extract all density-based clusters with respect to any distance ε ’ smaller than the generating distance ε from this order. 11 15 March 2010
14. Extract sub-trajectory clusters Scanning the order and assigning the cluster-membership Check the ascription of each sub-trajectory in cluster 12 15 March 2010
15. Density-based clustering OPTICS Motivation input parameters (e.g., Eps) are difficult to be determined We let the user to explore the clustering result at different scale one global parameter setting may not fit all the clusters Through the visualization it’s good to allow users to have flexibility in selecting clusters Yes! 13 15 March 2010
19. Conclusion A new framework to discover multi-granularity clusters by clustering sub-trajectories. An extension of the OPTIC algorithm for curve segments by using Fréchet distance measurement. The user can tune the density to explore the visualization of clustering results at different spatial granularities. 17 15 March 2010
20. Thank you for your attention. Questions? 18 15 March 2010
Hinweis der Redaktion
Morning everyoneNow I would like to give a brief introduction about my previous work, actually it`s my first time to attend this workshop and ICDM.So if you have any questions during my talk, pls feel free to ask meOk, then So our motivation of this work is to analyze ...The result of this research could be apply to …Next slide we will give two picture to explain this motivation more vividly
Let us take a look at the left picture , This is a part view of beijing planet map. The red line means the hot tourist path at this coarse level, Of course it dosent come from real data, we draw the line here just want to show our idea.At this level, one centimeter is equal to half a kilometer. Suppose we get the red line by clustering trajectories with some density measure, If we want to look more details of the Forbidden City, we should zoom in as much as is needed.Then we could get the right picture, one centimeter in the picture is equal to two hundred meters, we think the clusters density should be consistent with this grain size , with more low density measure we could get the cluster as this red line
The trajectory we mentioned
Let us give an illustration of the Frechet distance: a man is walking witha dog on a leash. This man is walking on the one curve, the dog on theother one. Both may vary their speed, but backtracking is not allowed.Then the Frechet distance of the curves is the minimal length of a leashthat is necessary. The Frechet method has the advantage of computing distancesonly on homologous points and not between closest points as for theHausdorff distance
The method at this time is not up to do this fast , the problem is to do this in a meaningful way