Coupled object detection and tracking from static cameras and moving vehicles.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multi-object tracking is crucial for understanding dynamic scenes.
• Existing methods often struggle with occlusions, detection errors, and lost tracks.

Purpose of the Study:

• To develop a novel, robust multi-object tracking approach.
• To couple object detection and spacetime trajectory estimation into a single optimization problem.

Main Methods:

• Formulated within a Minimum Description Length (MDL) hypothesis selection framework.
• Integrates state-of-the-art object detection (cars, pedestrians) with 3D scene geometry.
• Employs a trajectory estimation module for physically plausible spacetime paths.
• Utilizes model selection for global optimization of trajectories frame-by-frame.

Main Results:

• The approach recovers from detection mismatches and temporarily lost tracks.
• It generates physically plausible spacetime trajectories for multiple objects.
• Feedback mechanism improves future object detection.
• Demonstrated performance on surveillance and in-vehicle urban scenarios.

Conclusions:

• Coupling detection and trajectory estimation offers a robust solution for multi-object tracking.
• The MDL framework enables effective hypothesis selection and track recovery.
• The method shows promise for real-world applications in complex environments.