DBSCAN-Based Tracklet Association Annealer for Advanced Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Increasing demand for autonomous driving and smart video surveillance fuels advancements in deep neural network-based multi-object tracking.
• Inherent limitations of video cameras, particularly in occluded environments, present persistent challenges for accurate object tracking.

Purpose of the Study:

• To propose a novel density-based object tracking technique to address noise vulnerability and occlusion issues in multi-object tracking.
• To enhance the robustness, trajectory separation, and real-time processing capabilities of existing object tracking methods.

Main Methods:

• Redesigned a density-based object tracking technique based on the DBSCAN algorithm, known for its noise robustness and nonlinear clustering capabilities.
• Integrated the proposed method as a post-processor to existing trackers to improve overall performance.

Main Results:

• The proposed technique demonstrated improved performance across several key indices compared to existing tracking methods.
• Noise suppression capabilities led to a significant improvement of over 10% in tracking performance when used as a post-processor.
• The method effectively reduces the difficulty of trajectory separation and facilitates real-time processing.

Conclusions:

• The developed density-based object tracking technique offers a robust solution for challenges posed by noisy and occluded environments.
• This method holds significant potential for real-world applications in industrial settings, including pedestrian analysis and surveillance security systems.