Robust Event-Based Object Tracking Combining Correlation Filter and CNN Representation.

Area of Science:

• Computer Vision
• Robotics
• Sensor Technology

Background:

• Object tracking with event-based cameras (dynamic vision sensors/DVS) faces challenges like noise, rapid event-stream changes, complex backgrounds, and occlusion.
• Existing methods struggle with the unique data characteristics of DVS, limiting their application in high-speed environments.

Purpose of the Study:

• To develop a robust and efficient event-stream object tracking method for dynamic vision sensors.
• To improve tracking performance in complex scenes with occlusions and background clutter.
• To leverage deep learning features for enhanced appearance representation in event-based tracking.

Main Methods:

• Event-stream encoding using rate coding.
• Feature extraction from hierarchical convolutional layers of a pre-trained Convolutional Neural Network (CNN).
• Correlation filter mechanism for efficient object tracking.

Main Results:

• The proposed method demonstrates robust tracking performance in complicated scenes with noise, complex textures, and occlusions.
• The approach is resilient to variations in object scale, pose, and non-rigid deformations.
• The correlation filter integration ensures high-speed tracking capabilities.

Conclusions:

• The developed method effectively addresses key challenges in event-stream object tracking.
• This approach enhances the applicability of event-based vision sensors in fields like autonomous driving and robotics.
• The combination of CNN features and correlation filters offers a promising direction for high-speed vision applications.