Abstract
The smart transportation is of great importance to reduce the traffic and predict the most frequent path taken by the vehicles. One of the key technologies is trajectory analysis, which can be used for the study of ride sharing, traffic flow, abnormal detection, and more. In this thesis, we provide an empirical survey on the trajectory analysis for abnormal detection. The analysis has four stages. In the first stage, importing, cleaning and pre-processing of the GPS data will be initiated. In the second stage, different outliers detection methods will be used to find the best similarity measure and clustering method. In the third stage the validation and analysis of the different clustering methods with all similarity measures will be compared according to the internal and stability measure criteria. In the fourth stage, the method with the best output will be visualized on a digital map to see the abnormal behavior. We implemented different similarity measures, including Frechet distance, Dynamic Time Warping distance, Longest Common Subsequence distance, Hausdorff distance, Edit distance, and different clustering methods: K Means, CLARA (Clustering Large Applications), PAM (Partitioning Around Medoids), Hierarchical clustering, Model Based Clustering and FANNY (Fuzzy Analysis Clustering). We evaluated those methods using the HsinChu Bus trajectories,GeoLife dataset, Rio De Janeiro bus dataset and Android app Go Track dataset. Experimental results show that Frechet distance and Hausdorff distance with Hierarchical clustering provides the best result in terms of internal and stability measure criteria. The higher value of Silhouette width and Dunn Index clearly signifies that Frechet and Hausdorff distance when used with Hierarchical clustering give clusters with strong bonds and hence are able to find the abnormality in the trajectory data.