Layered Graphical Models for Tracking Partially Occluded Objects.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Occlusion by movable objects is a common challenge in computer vision systems, particularly for pedestrian tracking.
• Existing methods struggle to robustly handle dynamic occlusions and interactions between people and objects.

Purpose of the Study:

• To develop a novel representation for scenes with relocatable objects that cause partial occlusions of people.
• To create a system capable of accurately tracking pedestrians in environments with occluding objects like parked vehicles.

Main Methods:

• Formulated an occluder-centric representation using graphical model layers.
• Modeled pedestrian motion as a Markov process on activity zones.
• Aggregated image evidence in depth-ordered 2D observation regions, accounting for object occlusions.
• Enabled interaction between layers for inter-layer movement and shared image evidence.

Main Results:

• Demonstrated the formulation's effectiveness in tracking pedestrians near parked vehicles.
• Achieved favorable comparison against sprite-learning, deformable contour, and detector-based pedestrian trackers.
• Showcased robust handling of mutual overlap and occlusions in observation regions.

Conclusions:

• The proposed graphical model layer representation provides an effective solution for tracking people in scenes with dynamic occlusions.
• This approach offers improved performance and robustness compared to existing methods for pedestrian tracking in complex environments.