Motion segmentation using Ward’s hierarchical agglomerative clustering for crowd disaster risk mitigation

Abstract

Motion segmentation has gained increasing attention due to its significance for many public surveillance applications, such as behavior understanding and density estimation. Thus ensuring crowd risk mitigation and safety. Although numerous solutions have been put out to deal with this issue, it remains challenging due to factors like scene fluctuations, cluttered backgrounds, and severe occlusions. The present paper proposes a novel framework, spatio-angular-shape Ward linkage (SAS-WL), to segment human motion based on tracked trajectories features in crowd scenes. In the proposed SAS-WL, first the motion features are captured from the crowd video through the generalized-KLT (gKLT) keypoint tracker. The tracklets are then examined to keep only the significant and long-term tracklets. Then, spatio-angular-shape motion features are obtained from those long-term tracklets. Next, Ward’s linkage was utilized for performing motion pattern segmentation. In addition, we propose an improved Ward’s linkage method and integrate it with the obtained features to increase the quality of the clustering results. One of the significant advantages of the SAS-WL is its capability to address various densities in different crowd scenes. Moreover, the SAS-WL achieves accurate segmentation results. We validated our findings with videos of real-world crowd scenes. We introduce a novel dataset along with its manually annotated ground truth, comprising 21 diverse crowd videos. This dataset aims to enhance the assessment of our framework’s performance. Obtaining results on a diversity of challenging crowd scenarios demonstrates that this framework provides precise motion segmentation results.