SimpleTrack: Rethinking and Improving the JDE Approach for Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Joint Detection and Embedding (JDE) methods in multi-object tracking (MOT) often fail due to target loss or occlusion.
• Existing JDE methods fuse motion and appearance information into a data association matrix, which is vulnerable to temporary target disappearances.

Purpose of the Study:

• To propose a novel association matrix, the Embedding and GioU (EG) matrix, to enhance data association robustness in MOT.
• To develop an effective re-identification feature fusion tracker, SimpleTrack, and a new tracking strategy to address detection target loss.

Main Methods:

• Introduced the Embedding and GioU (EG) matrix, combining embedding cosine distance and Generalized Intersection over Union (GioU) distance.
• Developed SimpleTrack, a bottom-up fusion tracker for re-identity features, and a novel tracking strategy to mitigate detection loss.
• Evaluated the EG matrix by integrating it with five state-of-the-art JDE-based methods.

Main Results:

• Replacing existing association matrices with the EG matrix significantly improved IDF1, HOTA, and IDsw metrics across multiple JDE methods.
• The proposed methods increased tracking speed by approximately 20% without compromising accuracy.
• SimpleTrack demonstrated superior data association capabilities, achieving 61.6 HOTA and 76.3 IDF1 on the MOT17 test set at 23 FPS.

Conclusions:

• The EG matrix offers a robust and effective solution for improving data association in JDE-based MOT systems.
• SimpleTrack and the proposed tracking strategy effectively mitigate target loss and enhance overall tracking performance.
• The EG matrix provides significant performance gains and speed improvements, making it a valuable contribution to MOT research.