A Top–Bottom Clustering Algorithm Based on Crowd Trajectories for Small Group Classification

The study of crowd movement has recently become a popular research topic due to the increasing frequency of public safety issues. Compared with human evacuation experiments and drills, which may have personal safety risks and require a large number of volunteers, a simple and convenient computer simulation has become the mainstream research method. The computer simulation first needs to characterize small groups in the crowd to model the motion state of crowds for more accurate crowd modeling. In this paper, a top-bottom hierarchical clustering algorithm based on off-line crowd trajectories is proposed to provide small group information for crowd motion simulation. First, an unmanned aerial vehicle and a tracking technology are used to capture the pedestrian flow and extract the pedestrian trajectory. Second, a top-bottom hierarchical clustering strategy is proposed to divide the crowd into groups, which solves the problem of the difficulty of ascertaining small groups. This method solves the problem of automatically determining cluster centers by using the improved density peak clustering algorithm combined with a greedy algorithm. One factor of distinguishing small groups is improved by replacing the angle of the direction of motion with the distance difference, thus reducing the computational complexity. Specifically, the mean distance between trajectories based on the Euclidean distance is used for the top-level coarse-grained clustering; then, the improved Hausdorff mean distance is determined in the bottom-level fine-grained clustering. Third, the proposed algorithm is validated by classifying the groups of pedestrians in real videos. The experiments show that the proposed method is applicable and effective.