Deep Affinity Network for Multiple Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multiple Object Tracking (MOT) is crucial for video analysis, typically involving object detection and data association.
• Current data association methods rely on handcrafted features, limiting tracking performance.
• Deep learning has advanced object detection but not yet fully addressed data association challenges.

Purpose of the Study:

• To apply deep learning to the data association problem in Multiple Object Tracking (MOT).
• To develop an end-to-end framework that jointly models object appearances and their affinities.
• To improve the reliability and accuracy of online tracking systems.

Main Methods:

• Proposed the Deep Affinity Network (DAN) for end-to-end data association.
• DAN learns object features at multiple abstraction levels and computes affinities via exhaustive pairing.
• The method handles object appearance and disappearance, enabling deep association into previous frames.

Main Results:

• Achieved state-of-the-art performance on MOT15, MOT17, and UA-DETRAC benchmarks.
• Demonstrated superior tracking accuracy across twelve evaluation metrics.
• The developed DAN approach offers reliable online tracking capabilities.

Conclusions:

• Deep learning can significantly enhance data association for Multiple Object Tracking (MOT).
• The Deep Affinity Network (DAN) provides an effective end-to-end solution for complex tracking scenarios.
• The open-source implementation facilitates further research and application in computer vision.