An incremental DPMM-based method for trajectory clustering, modeling, and retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Trajectory analysis is crucial for video indexing, activity recognition, and surveillance.
• Existing methods often require predefined cluster numbers and struggle with dynamic data.

Purpose of the Study:

• To develop an adaptive and scalable trajectory clustering, modeling, and retrieval system.
• To automatically determine the optimal number of trajectory clusters.
• To enable online identification and incorporation of new trajectory clusters without retraining.

Main Methods:

• An incremental Dirichlet process mixture model (DPMM) for trajectory clustering.
• A time-sensitive Dirichlet process mixture model (tDPMM) for learning trajectory patterns and time-series characteristics.
• Development of a novel likelihood estimation algorithm for tDPMM and a trajectory-based video retrieval model.
• Combination of tDPMM probabilistic matching and DPMM model growing for scalability and adaptability.

Main Results:

• The proposed incremental DPMM effectively clusters trajectories and determines the number of clusters automatically.
• New trajectory clusters are identified and added online without prior data retraining.
• The tDPMM-based retrieval model demonstrates scalability and adaptability, outperforming state-of-the-art algorithms in experimental comparisons.

Conclusions:

• The developed DPMM and tDPMM framework provides an effective and adaptive solution for trajectory analysis.
• The system offers significant improvements in clustering, modeling, and retrieval of trajectory data, particularly in dynamic environments.
• This approach enhances the performance and scalability of motion event analysis in video applications.