Variational Inference for 3-D Localization and Tracking of Multiple Targets Using Multiple Cameras.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Multi-camera systems are crucial for 3-D localization and tracking.
• Challenges include back-projection uncertainty, ground point detection issues, occlusions, and unknown human heights.
• Existing methods struggle with complex, real-world scenarios.

Purpose of the Study:

• To propose a novel unified framework for 3-D localization and tracking in multi-camera settings.
• To address uncertainties caused by occlusions and unknown object heights.
• To enhance tracking accuracy and robustness in challenging environments.

Main Methods:

• Developed a Bayesian learning framework to maximize posterior probability over trajectory assignments and 3-D positions.
• Employed an expectation-maximization (EM) scheme with variational inference approximation.
• Utilized Boltzmann distributions derived from multi-camera tracking configurations.

Main Results:

• The proposed framework effectively handles uncertainties in multi-camera 3-D localization and tracking.
• Demonstrated superior performance compared to state-of-the-art methods on challenging datasets.
• Achieved accurate trajectory assignments and 3-D position estimations.

Conclusions:

• The novel Bayesian framework provides a robust solution for multi-camera 3-D localization and tracking.
• The expectation-maximization approach with variational inference offers a tractable method for complex problems.
• This work advances the state-of-the-art in multi-camera tracking for occluded and height-ambiguous scenarios.