Extendable Multiple Nodes Recurrent Tracking Framework With RTU+.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multiple-object tracking (MOT) commonly uses a tracking-by-detection paradigm.
• Current MOT methods often lose long-term tracking information due to detection failures or occlusions.
• Existing approaches rely on manual scoring functions for track proposals.

Purpose of the Study:

• To introduce a novel framework, Extendable Multiple Nodes Tracking (EMNT), for improved multiple-object tracking.
• To address the loss of long-term tracking information in current MOT methods.
• To develop a General Recurrent Tracking Unit (RTU++) capable of capturing long-term dependencies.

Main Methods:

• Developed the Extendable Multiple Nodes Tracking (EMNT) framework using four node types: correct, false, dummy, and termination.
• Proposed the General Recurrent Tracking Unit (RTU++) for scoring track proposals and capturing long-term information.
• Introduced an efficient method for generating simulated tracking data to augment limited datasets.

Main Results:

• Achieved state-of-the-art performance on MOT17, MOT20, and HiEve benchmarks.
• Demonstrated significant improvements when RTU++ was integrated into other trackers like MHT.
• Validated the generalization ability of RTU++ on MOTS20 and CTMC-v1 datasets using simulated data.

Conclusions:

• The EMNT framework and RTU++ effectively capture long-term information in multiple-object tracking.
• RTU++ shows strong performance and adaptability, even when trained on simulated data.
• The proposed methods offer a robust solution for various multi-object tracking scenarios.