Multi-View Multi-Human Association With Deep Assignment Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Accurate person identification across multiple camera views is crucial for various vision applications.
• Existing methods often struggle with more than two views or unknown camera perspectives.
• Handling scenarios where individuals are not visible in all views presents a significant challenge.

Purpose of the Study:

• To develop an end-to-end deep network framework for Multi-view Multi-Human Association (MvMHA).
• To address the complex problem of identifying individuals across multiple, non-fixed, and potentially overlapping camera views.
• To enable robust human association even when individuals are present in only a subset of views.

Main Methods:

• Utilizing an appearance-based deep network to extract subject features from each image.
• Computing pairwise similarity scores to construct a comprehensive affinity matrix.
• Employing a Deep Assignment Network (DAN) to generate an assignment matrix for MvMHA.

Main Results:

• The proposed framework successfully performs Multi-view Multi-Human Association in challenging scenarios.
• Effectiveness verified on both synthetic and real-world image datasets.
• Demonstrated strong cross-domain performance on three public datasets.

Conclusions:

• The developed deep network framework provides an effective solution for Multi-view Multi-Human Association.
• The approach is robust to variations in the number of views and individual visibility.
• The method shows promising generalization capabilities across different datasets.