Coherent crowd analysis with visual attributes / Nurul Japar

• 主要作者: Nurul , Japar
• 格式: Thesis
• 出版: 2022
• 主题: QA75 Electronic computers. Computer science

As human crowds become a norm due to the increasing global population, crowd analysis becomes essential to facilitate crowd surveillance. Towards improving surveillance tasks, extensive computer vision studies focus on analyzing coherent behavior in human crowds. Therefore, contextual information from visual attributes is essential in learning semantic relations among individuals. However, extracting discriminative visual attributes remains challenging due to challenges such as inter-object occlusions within crowd scenes. Hence, this thesis proposes solutions in analyzing coherent behavior in crowd scenes with visual attributes. This thesis first demonstrates the solution of exploiting contextual information to extract visual attributes within crowd scenes. Specifically, this thesis proposes a visual attributes extraction module to extract head-level visual attributes based on individual detection and head pose classification. Explicitly, it focuses on individuals’ head features to localize individuals and classify their head poses to distinguish individuals in crowds independently. Contrariwise to existing studies that focus on point-head annotations in {푥, 푦} coordinates, this module extracts visual attributes with spatial location, area of the bounding box, and head pose classification in {푥, 푦, 푤, ℎ} bounding boxes. Second, this thesis presents a coherent group detection framework to detect collective behavior in crowds by utilizing the visual attributes extraction module. Coherent groups represent individuals that are connected by collective behavior within crowd scenes. Unlike existing studies that focus on temporal information, the proposed framework detects the collective behavior by computing attributes similarity on individuals’ heads visual attributes. Via a clustering approach, the connected individuals are aggregated into local clusters for coherent group detection. These clusters represent mid-level representations of crowd understanding that illustrate group behavior. Third, this thesis extends the coherent group detection framework towards scene understanding. Specifically, a collectiveness analysis framework is designed to quantify and detect collectiveness from individual-level to scene level. The incremental learning in this framework notably analyzes semantic relations among individuals and infers topological relationship propagation via a manifold learning algorithm. Contrary to existing approaches, this approach computes crowd estimation for collectiveness quantification. It also computes the similarity and merges local clusters into global clusters for collectiveness detection. Inclusive experiments on various crowd scenes, i.e., Shanghai Tech RGB-D (ST RGB-D) Dataset, Collective Motion Database and CUHK Dataset, are conducted to demonstrate the efficacy of the proposed approaches. This thesis also presents several potential applications to facilitate crowd surveillance. As a result, the contributions of this thesis constitute more effective solutions for visual attributes extraction, coherent group detect and collectiveness analysis. Research findings from this thesis can assist as reference sources for the research community to support future work of crowd analysis.