Tracking Beyond Detection: Learning a Global Response Map for End-to-End Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing Multi-Object Tracking (MOT) methods often rely on a Tracking-by-Detection and Data Association paradigm.
• Deep neural networks have improved appearance features for cross-frame association, but the framework remains computationally intensive and not fully end-to-end.
• The Tracking-by-Detection approach faces limitations in inference speed and overall performance.

Purpose of the Study:

• To develop an effective end-to-end deep learning framework for Multi-Object Tracking (MOT).
• To enable direct processing of image sequences/videos for object localization and tracking.
• To overcome the computational and performance limitations of existing MOT frameworks.

Main Methods:

• A novel global response network is proposed to project objects into a continuous response map.
• Object trajectories are extracted directly from the generated response map.
• The framework takes image sequences/videos as direct input, outputting located and tracked objects.

Main Results:

• The proposed online tracker achieves state-of-the-art performance on the MOT16 and MOT17 benchmarks.
• Experimental results demonstrate significant improvements in tracking metrics.
• The end-to-end deep learning framework shows high effectiveness in Multi-Object Tracking.

Conclusions:

• The developed end-to-end deep learning framework offers an effective solution for Multi-Object Tracking.
• The novel global response network approach simplifies trajectory extraction and improves efficiency.
• The method achieves state-of-the-art results, highlighting its potential for real-world applications.