Multi-objective pedestrian tracking method based on YOLOv8 and improved DeepSORT.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian tracking in complex traffic scenarios faces challenges like occlusion and ID switching.
• Existing methods often struggle with dynamic target re-identification and trajectory prediction accuracy.

Purpose of the Study:

• To develop a robust multi-objective pedestrian tracking method (YOFGD) addressing occlusion and ID transformation in real-world traffic.
• To enhance the accuracy and real-time performance of pedestrian detection and tracking systems.

Main Methods:

• Utilized You Only Look Once version 8 (YOLOv8) for enhanced feature extraction and Omni-Scale Network (OSNet) for real-time tracking.
• Integrated an adaptive forgetting smoothing Kalman filtering algorithm (FSA) and Complete-Intersection over Union (CIOU) for improved data association.
• Developed a Generalized Trajectory Feature Extractor Model (GFModel) to merge local and global trajectory information.

Main Results:

• The YOFGD method achieved a tracking accuracy of 77.9%.
• The system demonstrated a real-time processing speed of 55.8 frames per second (FPS).
• Significant improvements were observed in reducing pedestrian omission and misdetection rates.

Conclusions:

• YOFGD effectively handles challenges in complex traffic scenarios, offering high accuracy and real-time performance.
• The proposed fusion of YOLOv8, improved DeepSORT, OSNet, FSA, and GFModel provides a robust solution for pedestrian tracking.
• The system's performance meets the demands of practical real-world applications.