The proposed segmentation and similarity assessment procedures rely on a set of thresholds that are required to be specified by the user prior to the analyses. These thresholds include:
• sinuosity threshold: to distinguish the low sinuosity from the high sinuosity segments in movement parameter profiles.
• deviation threshold: to distinguish the points with low versus high ampli- tude levels along the movement profile in the segmentation process. • distance threshold: to determine similar entities in the similarity assessment
The determination of these thresholds depends on a variety of parameters such the spatial and temporal scale of the processes under study, spatial and temporal granularity (i.e. sampling interval), and the noise level of the movement data. Moreover, the purpose of similarity assessment and trajectory clustering affects the choice of the required thresholds. Therefore, a sensitivity analysis is sug- gested as a prerequisite of running the proposed methods, in order to adapt the thresholds to the subject data.
The effects of outliers
The segmentation process relies on the amplitude level of movement parameters. Therefore, the presence of outliers in the data affects the accuracy of the results. Hence, as described in Research Paper 2, a preprocessing step including trajectory filtering is required prior to the segmentation of the trajectories.
Limitations of the Trajectory Classification
The proposed trajectory classification using the segmentation technique (see Re- search Paper 2) relies on the distinctions between the movement characteristics (e.g. speed, acceleration etc.) of the various objects. However, the effective- ness of this method decreases when different types of objects exhibit similar movement characteristics. For instance, in the transport mode detection in very heavy traffic, different types of vehicles to a great degree exhibit similar move- ment characteristics (e.g. very slow movement). Therefore, it becomes difficult to extract the mode of transport of vehicles from movement data alone. This is also observed on the other for other transport mode detection techniques (Upadhyay et al., 2008).
Movement path similarity
The proposed similarity assessment technique determines the similarity between the movement characteristics of objects in terms of the variations of movement parameters. In this thesis, the similarity of the movement path generated by the dynamic objects was not the main focus of the methods. It was, however, investigated implicitly in the form of coincidence movement patterns in Research Paper 4. The available spatial similarity techniques such as the ones introduced in section 2.5.4 could be applied for this purpose.
Efficiency and Scalability
The edit-distance based similarity analysis approaches, including our proposed NWED measure, have a relatively high computational cost (i.e. O(n2), where
n is the number of observations along a trajectory). Therefore, the efficiency of this measure decreases for very long trajectories in large movement datasets. Although the segmentation process improves the storage and pattern matching of
the trajectories, additional strategies should be considered to improve the com- putations of similarity in such datasets. In order to overcome this weakness, applying indexing, dimension reduction, and pruning approaches would be bene- ficial (Vlachos et al., 2002a, 2004; Chen et al., 2004, 2005; Ding et al., 2008b).
Conclusions
This thesis presented research on the development of methods for knowledge dis- covery from movement data about dynamic objects and processes. The main motivation was to exploit the movement parameters (e.g. speed, acceleration, turning angle) in support of the study of the collective movement behavior of ob- jects. Hence, the main focus was on the dynamic behavior of objects rather than the geometric specifications of the objects’ lifelines (e.g. geometry of the move- ment paths). The general aim of this chapter is to highlight the most pertinent outcomes of this thesis and draw attention to some outlooks on future work.
5.1 Main Contributions
In response to the research objectives, this thesis sought to develop conceptual and methodological knowledge about movement parameters and their features. This research mainly contributes to knowledge discovery from movement data with the development of a quantitative approach that enables similarity assess- ment of the movements of dynamic objects. The methodology uses the movement parameters that can be either derived from the trajectory of objects or directly recorded by new generations of mobile sensor technologies.
This thesis pursued a three-stage research process towards the main goal. The achievements of each stage account for a portion of the contributions made in this thesis. Accordingly, this thesis brought about the following main achievements: E Stage I developed a conceptual framework of movement. The framework
encompasses the fundamental elements of the movement of objects. These elements are required in the identification and formalization of movement pat- terns. The conceptual framework formed the basis for the subsequent stages. Furthermore, the first stage of this research introduced a comprehensive clas- sification of movement patterns. The proposed conceptual framework and classification should facilitate the development of pattern recognition algo- rithms that are required to be efficient, effective, and as generic as possible in a more systematic approach.
E Stage II developed a feature extraction and segmentation method to reduce the complexity of moving object trajectories for the purpose of analysis and knowledge discovery. In this thesis a segmentation technique was conceived
to generate a concise view of trajectories encapsulating the important move- ment features. Segmentation aims at decomposing data into parts of similar characteristics. Hence, the developed method can serve to assess the simi- larity of the movement characteristics of multiple objects in the classification (e.g. transport mode detection) and clustering of movement data. The de- veloped segmentation algorithm forms the key element of the main movement similarity assessment approach in the subsequent stage.
E Stage III proposed two different similarity assessment approaches to extract the similarities between the movement characteristics of objects. Both meth- ods detect objects whose movement parameters exhibit similar patterns. The main method applies the segmentation process, proposed in Stage II, on a sin- gle movement parameter profile. The NWED distance was introduced, based on the edit distance, as the similarity measure of the segmented profiles. The alternative, second method assesses the similarity of dynamic objects in a multidimensional movement parameter space using the average Euclidean dis- tance. Furthermore, the final stage introduced clustering and movement pat- tern extraction strategies relying on the above similarity assessment methods. The applicability of the developed techniques was evaluated on clustering and movement pattern discovery from the movements of dynamic objects. It is necessary to remark that this thesis investigated the applicability of the pro- posed methods on real movement datasets from different application domains. Sample data included, movement data from meteorology (i.e. hurricanes), differ- ent data sources in transportation (i.e. pedestrians, bicycles, motorcycles, cars), and eye-tracking. The experimental results of this thesis indicate that the pro- posed methods could be successfully applied in support of movement behavior studies of dynamic objects and processes. The methods were developed generi- cally so that they could be applied to any kind of movement data from various application domains such as movement ecology of animals, urban ecology, and human mobility studies.