High-Speed Multiple Object Tracking Based on Fusion of Intelligent and Real-Time Image Processing

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Multiple object tracking (MOT) is crucial for applications like autonomous driving and surveillance.
• Existing MOT methods often prioritize association over detection speed, limiting real-time performance.
• There's a need for MOT systems that balance speed, accuracy, and robustness.

Purpose of the Study:

• To develop a high-speed MOT system that enhances real-time performance without sacrificing tracking accuracy.
• To investigate the impact of accelerating detection on overall MOT system efficiency.
• To propose a novel hybrid tracking framework and tracker management strategy.

Main Methods:

• A hybrid tracking framework combining low-frequency deep learning detection with classical high-speed tracking.
• A detection label-based strategy for managing object tracks.
• Evaluation in six scenarios using high-speed camera data and comparison with seven state-of-the-art (SOTA) methods.

Main Results:

• Achieved high frame rates: up to 470 fps (2 objects), 243 fps (3 objects), and 178 fps (4 objects).
• Secured top scores in MOTA, IDF1, and HOTA with high-accuracy detection.
• Demonstrated effective long-term association for high-speed tracking, even with lower detection accuracy.

Conclusions:

• The proposed system offers a practical and efficient baseline for high-speed MOT.
• The multi-processing architecture advances MOT research, particularly for systems with asynchronous modules.
• The hybrid approach effectively balances real-time performance, tracking accuracy, and robustness.