Siam Deep Feature KCF Method and Experimental Study for Pedestrian Tracking.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Pedestrian tracking is crucial for public safety, especially for unmanned aerial vehicle (UAV) systems.
• Traditional Correlation Filter (CF) algorithms struggle with fixed template sizes and occlusions, limiting their effectiveness in real-world scenarios.

Purpose of the Study:

• To develop a robust and accurate pedestrian tracking framework for UAVs that overcomes the limitations of existing methods.
• To enhance tracking performance, particularly in scenarios involving occlusions and re-identification.

Main Methods:

• A novel tracking-by-detection framework integrating a lightweight YOLOv3 (You Only Look Once version 3) with Efficient Channel Attention (ECA) for deep feature extraction.
• Incorporation of a lightweight Siamese Convolutional Neural Network (CNN) with Cross Stage Partial (CSP) for robust feature representation and target similarity in data association.
• Development of the Deep Feature Kernelized Correlation Filters (DFKCF) method coupled with Siamese-CSP (Siam-DFKCF).

Main Results:

• The proposed Siam-DFKCF method demonstrated significantly improved anti-occlusion and re-tracking performance compared to existing approaches.
• Experimental results showed increased tracking accuracy, with Distance Precision (DP) reaching 0.934 and Overlap Precision (OP) reaching 0.909 on test data.

Conclusions:

• The Siam-DFKCF framework offers a robust solution for pedestrian tracking in computer vision applications, particularly for UAV systems.
• The integration of deep learning features and advanced correlation filtering techniques enhances tracking reliability and accuracy, especially under challenging conditions like occlusions.