Greedy Batch-Based Minimum-Cost Flows for Tracking Multiple Objects.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Algorithm Design

Background:

• Minimum-cost flow (MCF) algorithms excel in multi-object tracking but require full image sequences.
• Real-time and distributed tracking often use batch processing, leading to error propagation due to decoupled tracking and stitching.
• Existing batch-based methods struggle with correcting errors across batch boundaries.

Purpose of the Study:

• To develop a greedy batch-based minimum-cost flow approach for robust multi-object tracking.
• To address the limitations of sequential batch processing in current tracking systems.
• To improve the accuracy and reliability of tracking in real-time and distributed environments.

Main Methods:

• A generalized minimum-cost flows (MCF) algorithm is applied to individual batches to generate probable and low-probability trajectories.
• Consecutive batches are jointly optimized, allowing tracking results from one batch to inform and correct another.
• A second application of the generalized MCF algorithm establishes optimal matching between trajectories across batches.

Main Results:

• The proposed approach effectively integrates tracking across batch boundaries, mitigating error propagation.
• Joint optimization of consecutive batches leads to more accurate and complete object trajectories.
• The method demonstrates strong performance across diverse datasets without requiring model training.

Conclusions:

• The greedy batch-based MCF approach offers a simple yet effective solution for multi-object tracking.
• Joint optimization of batches significantly enhances tracking accuracy compared to sequential methods.
• This method provides a robust and adaptable solution for real-time and distributed tracking scenarios.