A Novel Vehicle Tracking ID Switches Algorithm for Driving Recording Sensors

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Real-time vehicle tracking is crucial for autonomous systems.
• Occlusion events frequently cause tracking errors, leading to ID switches.
• Existing methods struggle with maintaining consistent vehicle identities during occlusions.

Purpose of the Study:

• To develop a novel method for robust vehicle tracking across video frames.
• To improve the accuracy of vehicle re-identification after occlusion.
• To reduce the occurrence of ID switches in real-time tracking systems.

Main Methods:

• Utilized YOLO v3 for object detection and a modified Deep SORT algorithm for tracking.
• Integrated a Kalman filter for vehicle position prediction.
• Incorporated Mahalanobis, cosine, and Euclidean distances, with appearance metrics enhancing cosine distance calculations.

Main Results:

• Achieved a 29.95% reduction in ID switches on the BDD100K dataset.
• Demonstrated a 32.16% reduction in ID switches on the COCO dataset.
• The proposed method effectively handles vehicle occlusion, maintaining tracking consistency.

Conclusions:

• The novel vehicle tracking method significantly enhances tracking robustness.
• Reduced ID switches improve the reliability of real-time vehicle monitoring systems.
• The approach shows promise for applications in autonomous driving and traffic analysis.