In this chapter a variety of knowledge discovery methods for large-scale 2D and 3D trajectory databases to efficiently incorporate existing prior knowledge are depicted in detail. Using these analysis methods tailored for tracking ap- plications, underlying phenomena and occurring effects can be investigated on a detailed single trajectory level allowing to efficiently gain knowledge about the trajectory databases. The conceptual approaches developed in this chapter are also applicable to fragmented tracking data allowing to incorporate even highly complex problem classes in the analysis. To efficiently incorporate ex- isting prior knowledge in the analysis of large-scale trajectory databases a new knowledge discovery process is introduced allowing to handle different origins of expert-tailored prior knowledge. Therefore, a wide variety of trajectory char- acteristics are introduced in this chapter depicting different aspects of prior de- fined trajectory characteristics. However, it is not feasible to automatically in- corporate not formalized prior knowledge yet. To cope with such new emerg- ing prior knowledge for application-specific trajectory datasets, a systematic procedure to efficiently incorporate new characteristics using the developed interface is described in this chapter. To efficiently handle the circumstance of error-prone fragmented tracking data an approach using representative tra- jectories resulting from initial automatic fusion routines or by region-specific manual curation effort was developed. This approach allows to allocate tra- jectory fragments to the predominant group frequency of the identified spatio- temporally nearest neighbors with a minimal effort of time. For spatially well separated groups the allocation approach provides good results. However, in highly dense regions with multiple groups intersecting especially in border re- gions, the allocation process is not suited well to perfectly allocate the remain- ing trajectory fragments. An interactive selection of representative trajectories located in the transition regions however can improve the allocation result, still allowing to make reasonable global assumptions about the fragmented trajec- tory dataset. To further analyze the trajectories in specific regions in detail, a feasible approach is introduced to globally truncate the trajectories used for analysis tasks to a fixed spatial region by specifying the time interval of interest allowing to evade to average out specific phenomena and behaviors of interest. Moreover, a feature-based group extraction approach suited for trajectory data
allows finding groups of possible interest by efficiently incorporating prevalent prior knowledge within clustering and classification task. However, not in all application-specific trajectory databases the groups of potential interest can be automatically extracted without the presence of convenient prior knowledge. To cope with the occurrence of sparse prior knowledge, Chapter 4 provides a variety of interactive visual representations to efficiently gain new knowledge and find underlying phenomena in large-scale trajectory databases. Once, rele- vant groups of potential interest are extracted within one 3D+t database using the knowledge discovery process developed in this chapter, the results can be efficiently transferred to new databases allowing the additionally handling of emerging inhomogeneities. However, an automatic transfer to new database comprising strong inhomogeneities is not possible, instead application tailored manual interactive adaptions are used to efficiently cope with this problem. Moreover, this chapter provides a wide range of methods to quantitatively de- scribe extracted groups within trajectory databases, allowing to make profound assumptions of inter- and intra-group characteristics leading to the discovery of phenomena that are not ostensible. To additionally access trajectory databases with an enormous complexity containing dividing object, this chapter offers methods for a quantitative lineage analysis comprising spatial and temporal aspects. These methods, however require the preceding curation of fragmented trajectory data to obtain valid assumptions about the underlying dividing ob- ject characteristics. Moreover, a systematic approach to classify a trajectory in the context of the surrounding neighborhood is depicted in this chapter allow- ing to make statements about the homogeneity of the embedded environment of each trajectory. In future work, trajectory group pattern such as flock and convoy behavior can be integrated in the existing interface provided by the actual knowledge discovery framework to complement the trajectory features catalogue. Furthermore, approaches to automatically extract relevant groups of potential interest without requiring existing prior knowledge and human effort are the challenging tasks for the future.
Prior Knowledge Integration in
Trajectory
Knowledge
Discov-
ery
The systematical integration of prior knowledge in the knowledge discovery process of large-scale trajectory databases was depicted in the previous chap- ter. Here, in Chapter 4, visual interactive approaches are presented for the efficient guidance of the knowledge discovery process and the correspond- ing prior knowledge allocation reaching from rough group assignments up to single object-tailored characteristics. The methods presented in Chapter 3 mainly focus on the automated knowledge discovery process in 3D+t trajec- tory databases, whereas Chapter 4 introduces interactive possibilities to incor- porate existing prior knowledge in the trajectory knowledge discovery process (Figure 3.3). Several visual representations of trajectory data are presented in Section 4.1 for a detailed investigation from multiple perspectives allowing to focus on a wide variety of attributes. Therefore, an efficient 3D representation of trajectories is introduced in Section 4.1.1. Furthermore, a maximum intensity projection overlay (Section 4.1.2), an efficient 3D+t visualization of migrating objects (Section 4.1.3) as well as a feature-based visual trajectory representation (Section 4.1.4) assure full power in interactively accessing trajectory databases. The visual handling of dividing object characteristics is described in Section 4.1.5. Furthermore, the incorporation of additional visual representation in the modular concept is depicted in Section 4.1.6. To interactively dissect the ex- isting trajectories and directly allocate the predefined prior knowledge to the intended location within the spatio-temporal paths of the objects, Section 4.2.1 yields a variety of approaches including online propagation of performed se- lections (Section 4.2.2). For an efficient visually guided hierarchical analysis,
Section 4.3 depicts a newly developed visualization possibility allowing the ex- pert to easily derive hypotheses and analyze phenomena of highly complex trajectory databases. In case of fragmented tracking data, Section 4.4 addition- ally provides the possibility to interactively curate tracking errors in an effi- cient way using a visually guided approach. The combination of the proposed visual representations tailored for large-scale 3D+t trajectory data combined with the possibilities of multi-linked selection strategies is new and pave the way to new approaches in the field of trajectory analysis (Chapter 7). The sin- gle visual representation themselves are not new from scratch. However, the ability to online propagate selection results between all mentioned visualiza- tions is new, allowing to incorporate prior knowledge from several sources in the trajectory knowledge discovery process. Furthermore, the visualizations are implemented in the way (Chapter 5) that analysis of large-scale trajectory databases incorporating hundreds of thousands of trajectories gets applicable. The evaluation of the methods presented in this chapter is only useful in combi- nation with the knowledge-discovery possibilities of Chapter 3 and the imple- mentation of the whole framework depicted in Chapter 5. Therefore, Chapter 6 presents a user study to evaluate the overall possibilities and interactive ap- proaches in combination.