Real-Time Robust 2.5D Stereo Multi-Object Tracking with Lightweight Stereo Matching Algorithm

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Monocular vision systems struggle with object tracking due to occlusions and truncations.
• Existing stereo vision methods for tracking are computationally intensive, requiring dense matching or 3D reconstruction.
• There is a need for efficient and accurate stereo multi-object tracking solutions.

Purpose of the Study:

• To develop a real-time 2.5D stereo multi-object tracking framework.
• To improve tracking robustness against occlusions using geometric constraints.
• To provide a computationally efficient alternative to existing stereo tracking methods.

Main Methods:

• Implemented a lightweight stereo matching module using Direct Linear Transform-based triangulation with bounding box coordinates.
• Utilized a dual-tracker architecture for independent view tracking and re-identification.
• Employed geometric constraints for robust correspondence without appearance features.

Main Results:

• StereoSORT achieved a Multiple Object Tracking Accuracy (MOTA) of 0.932 and an Identification F1 score (IDF1) of 0.823.
• Outperformed monocular trackers like OC-SORT (IDF1: 0.765) and ByteTrack (IDF1: 0.609).
• Achieved a median depth error of 50.1 mm and a processing speed of 70 FPS.

Conclusions:

• Geometric constraints alone are sufficient for robust stereo tracking, eliminating the need for costly appearance features.
• StereoSORT offers a practical and efficient solution for multi-object tracking in resource-constrained environments.
• The framework demonstrates high accuracy and reliability comparable to commercial sensors.