Mixture of Switching Linear Dynamics to Discover Behavior Patterns in Object Tracks

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object tracking often requires understanding both low-level motion dynamics and high-level behavioral patterns.
• Existing models may struggle with complex, variable motion data common in surveillance and autonomous systems.

Purpose of the Study:

• To develop a novel non-parametric Bayesian model for joint discovery of object actions and behaviors.
• To handle real-valued features directly, avoiding information loss from quantization.
• To accurately model complex kinematic and spatial variations in object tracks.

Main Methods:

• A non-parametric Bayesian model is proposed, integrating actions (linear dynamics with spatial distribution) and behaviors (Markov chains of actions).
• Dirichlet Processes are used to automatically discover the number of actions and behaviors from data.
• Inference is performed using Markov Chain Monte Carlo (MCMC) sampling.

Main Results:

• The model successfully distinguishes relevant behavior patterns in pedestrian tracks, outperforming state-of-the-art hierarchical and simpler models.
• Validation on artificial and real-world datasets from surveillance and intelligent vehicles demonstrates robustness.
• The approach handles real-valued features, enabling nuanced discovery of variations like walking and running without discrete thresholds.

Conclusions:

• The novel Bayesian model effectively captures complex object dynamics and behaviors.
• This method offers improved performance for analyzing object tracks in challenging environments.
• Public release of software and datasets facilitates future research in behavior recognition.