Achieving Adaptive Visual Multi-Object Tracking with Unscented Kalman Filter.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multi-object tracking is crucial for intelligent systems but faces challenges like occlusion and motion changes.
• Existing algorithms struggle with accuracy and robustness in complex, dynamic scenes.

Purpose of the Study:

• To enhance the speed and accuracy of multi-object tracking.
• To address challenges in occlusion, background noise, and rapid motion changes.
• To improve the robustness of tracking algorithms in complex environments.

Main Methods:

• Developed an improved DeepSORT algorithm integrated with YOLOv5 for object detection.
• Implemented a general object motion model and an Unscented Kalman Filter (UKF) for motion state estimation.
• Introduced an adaptive factor to manage observation noise and adjust the innovation covariance matrix.

Main Results:

• The improved DeepSORT algorithm demonstrated a 4.75% increase in speed and a 2.30% increase in precision compared to the original DeepSORT.
• Achieved superior performance, particularly on the MOT16 dataset with dynamic cameras.
• Showcased enhanced robustness and accuracy in complex occlusion scenarios.

Conclusions:

• The proposed YOLOv5-based improved DeepSORT algorithm significantly enhances multi-object tracking performance.
• The algorithm offers a robust solution for real-world applications requiring accurate and fast tracking.
• Further improvements in tracking accuracy and robustness were validated through extensive experiments.