Extracting Vehicle Trajectory Clusters from Video Monitoring in Smart Intersections

The paper investigates clustering of vehicle trajectories crossing through an intersection monitored with an AI-based video monitoring system, in order to estimate the amount of traffic passing from each input to each output street of the intersection. We show that street-to-street traffic can be obtained using distance-based trajectory clustering. We evaluate several distance measures and evaluation scores for trajectory clustering in the context of our system, and report the results. We find that, in our setup, the Hausdorff distance together with the Silhouette score provide the best clustering results.