Modeling and classifying human activities from trajectories using a class of space-varying parametric motion fields

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Trajectory analysis often uses probabilistic generative models, but struggles with space-dependent dynamics.
• Switched linear dynamical models (e.g., Hidden Markov Models) capture motion regimes but not spatial variations.
• Nonlinear models can capture space-dependent dynamics but are complex to identify.

Purpose of the Study:

• To propose a new trajectory modeling approach using a mixture of parametric motion vector fields.
• To enable the representation of diverse trajectories and space-dependent behaviors.
• To address limitations of existing models in capturing complex motion dynamics.

Main Methods:

• Modeling trajectories with a mixture of parametric motion vector fields.
• Employing a probabilistic mechanism with a field of stochastic matrices for switching between fields.
• Avoiding global nonlinear dynamical models for simpler identification.

Main Results:

• The proposed method effectively models a wide variety of trajectories.
• Demonstrated capability in capturing space-dependent behaviors.
• Successful experimental evaluation on synthetic and real-world human trajectory data.

Conclusions:

• The mixture of parametric motion vector fields offers a flexible and effective approach to trajectory modeling.
• This method overcomes limitations of linear models in handling space-dependent dynamics.
• The approach shows promise for applications like human activity classification in video surveillance.