Cross-scale analysis in classification and segmentation of movement

Abstract

Movement is the essence of many spatiotemporal phenomena around us. Recent advances in tracking technologies have enabled the collection of tremendous amounts of movement trajectory data. Following in the footsteps of data production, computational methods are being developed in order to extract meaningful patterns from the raw movement data. These patterns, in return, can be related to valuable information about the behaviors of the moving objects under study. However, due to the internal and external factors influencing movement, the behaviors maybe compounds of different patterns at various spatial and temporal scales. The focus of this thesis therefore lies on investigating the importance of scale and cross-scale analysis in two movement analysis tasks, namely movement classification and trajectory segmentation. In movement classification, the aim is to build a classification model by finding relationships and rules among movement features in order to assign the input data to known classes. In trajectory segmentation, however, the aim is to decompose a movement trajectory into segments of homogenous movement characteristics. These characteristics can be measured by different geometrical, physiological or semantic properties of movement. The relevance of these two analysis tasks are highly recognized in the literature, however, the consideration of cross-scale aspects has the advantage to improve the commonly used single-scale approaches in the tasks of movement classification and trajectory segmentation. The main contribution of this thesis lies in introducing new methodologies for cross-scale movement analysis. In movement classification, we employed a resampling method for an improvement computation of movement parameters across different temporal scales as input features in the classification. Moreover, the use of discrete wavelet transform (DWT), as another multi-scale measure, is investigated to provide complementary features in movement classification. DWT is further used in trajectory segmentation, where the provided decomposition levels of DWT is used to investigate the variations in movement patterns across different scales. In the addressed tasks, this thesis shows that cross-scale analysis is needed in order to define an analysis scale which matches better to the scale of phenomena under study and that employing such methods yields better-quality results compared to single-scale analysis. The importance of cross-scale analysis was revealed by application on various movement datasets in real-world applications such as neuropharmacology, behavioral ecology, and biology.