Learning Trajectory Patterns by Clustering: Comparative Evaluation
Problem Description & Definition
Scene Modeling -Clustering Approaches
Distance/Similarity Measures Euclidean/Modified Euclidean distance Dynamic Time Warping Longest Common Sequences modified Hausdorff Distance 1/15/2018
Clustering Methods Iteration optimization Online adaptive K-Means, Fuzzy C-Means Online adaptive incremental learning Hierarchical clustering agglomeratvie & divisive, min-cut graph based, dominant set clustering Co-occurrence de-compostion document and keywords Note: For clustering methods we decide to utilize CLUTO[7] data clustering software package, which is used for agglomerative, divisive, hybrid and graph-based clustering. The software provides a number of options and optimization criteria for each cluster variants. In matlab K-Means clustering is available to use directly. 1/15/2018
Evaluation(Dataset) We decided to use Lankershim dataset (a total of 30 minutes vehicle trajectories processed from the video data) provided by the Federal Highway Administration (FHWA) and the Next Generation Simulation (NGSIM). 1/15/2018
Evaluation(Criteria) We will manually label the trajectories into cluster and evaluate different distance metrics and clustering algorithms in terms of correct clustering rate (CCR) where N is the total number of trajectories and pc denotes the total number of trajectories matched to the c-th cluster 1/15/2018