A Comparative Study of Features and Distance Metrics for Trajectory Clustering in Open Video Domains

Spatio-temporal trajectory is one of the most important features for understanding activities of objects. Trajectory clustering can thus be used to discover different motion patterns and recognize event occurrences in videos. Similarity measure plays the key role in trajectory clustering. In this paper, we conduct a comparative study on different features and distance metrics for measuring similarities of trajectories from open video domains. The features include the location of each point on the trajectory, velocity and direction (curvature and angle) of motion along the timeline. The distance metrics include Euclidean distance, DTW (Dynamic Time Warping), LCSS (Longest Common Subsequence), and Hausdorff distance. Besides, we also investigate the combination of different features for trajectory similarity measure. In our experiments, we compare the performances of different approaches in clustering trajectories with various lengths and cluster numbers.