Measuring the effectiveness of various trajectory similarity measures

The measurement of trajectory similarity plays a critical role in analyzing and interpreting trajectory data generated by vehicles, individuals, and other moving entities. This thesis conducts a comprehensive evaluation of various trajectory similarity measures across a series of clustering and retrieval experiments. Both synthetic and real-world datasets of varying characteristics were employed to ensure diverse evaluation scenarios. The methodology involved comparing each measure in terms of execution time, memory consumption, and output quality. For clustering tasks, hierarchical clustering with average linkage was applied, and cluster quality was assessed using validity indices and visual inspection. For retrieval tasks, a progressive query reduction strategy was used to examine each similarity measure’s ability to recognize similar trajectories under decreasing levels of information. The results revealed consistent patterns: Jaccard, DTW, LCS, and CAE were the fastest in execution, though the latter three exhibited higher memory consumption. DTW, LCS, and Jaccard also consistently produced high-quality clustering and retrieval results, with LCS and Jaccard excelling due to their parameter customization capabilities. In contrast, measures such as DFD, ERP, EDR, OWD, and MSM were considerably slower and less effective. Additionally, implementation differences significantly affected performance, indicating that efficiency may vary across different software libraries or future improvements. Ultimately, this work identifies Jaccard, LCS, and DTW as the most reliable similarity measures within the evaluated context, offering a foundation for future studies to explore scalability, newer implementations, and broader real-world applications.