Contextualized trajectory parsing with spatiotemporal graph.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Automated analysis of surveillance video is crucial for security and monitoring.
• Parsing object trajectories requires addressing spatial segmentation, temporal tracking, and object categorization simultaneously.

Purpose of the Study:

• To develop a unified framework for automatic object trajectory parsing in surveillance videos.
• To introduce a novel spatiotemporal graph (ST-Graph) representation for motion primitives.

Main Methods:

• Proposed a spatiotemporal graph (ST-Graph) where nodes represent motion primitives and edges denote relationships between them.
• Formulated trajectory parsing as a graph multicoloring problem with a unified probabilistic approach.
• Employed an efficient composite cluster sampling algorithm for optimal solution search.

Main Results:

• The ST-Graph effectively integrates context knowledge as informative priors.
• The proposed framework demonstrated state-of-the-art tracking accuracy on challenging public datasets.
• Successfully addressed joint spatial segmentation, temporal tracking, and object categorization.

Conclusions:

• The novel ST-Graph representation and probabilistic formulation offer an effective solution for complex trajectory parsing tasks.
• The developed framework advances the state-of-the-art in automated surveillance video analysis.
• This approach provides a robust method for understanding object motion patterns in video data.