Motion segmentation of pedestrian trajectories Using Angular Gaussian Mixture model

Abstract

Motion segmentation and identification of structural segments from an ensemble of human trajectories continue to be a challenge. These processes entail distinguishing and categorizing distinct motion patterns manifested by pedestrians within a group, as well as recognizing conspicuous segments that exhibit their overall motion. Accurate motion segmentation is vital in many areas, particularly computer vision, automation, and human behavior investigation and monitoring. However, owing to the complexity and irregularity of human motion, this task demands sophisticated procedures. Addressing this challenge, the Angular Gaussian Mixture Model (AGMM) is proposed in this study for visual motion segmentation. The angular features of the pedestrian trajectories are incorporated into the GMM, allowing the overall proposed framework to accurately expose the similarity between the trajectories and fulfill motion segmentation. The experimental findings conducted on the CUHK benchmark demonstrate that the proposed framework outperforms various state-of-the-art methods. The findings signify the superior performance achieved by the proposed approach in effectively and accurately segmenting motion trajectories within different crowded scenarios. Statistical evaluation approaches, such as normalized mutual information index (NMI), purity, rand index (RI), F-measure, also known as the F1-score (F1), and accuracy (ACC), were utilized employing an immense number of real-world video clips. This study has effectively laid the foundation for an essential and initial stride towards achieving comprehensive, high-level crowd behavior analysis.