Distractor-Aware Event-Based Tracking.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Event cameras, or dynamic vision sensors, offer advantages in capturing dynamic scenes due to asynchronous intensity change detection.
• Existing event-based trackers often adapt RGB methods, limiting robustness in challenging scenarios like low light or fast motion.
• Conventional trackers struggle with cluttered backgrounds and camera motion, necessitating new approaches.

Purpose of the Study:

• To develop a distractor-aware event-based tracker (DANet) that enhances object tracking performance.
• To exploit both motion cues and object contours from event data for improved target distinction.
• To create a robust tracker capable of handling dynamic distractors and challenging environmental conditions.

Main Methods:

• Proposed a Siamese network architecture incorporating transformer modules, named DANet.
• Developed a dual-network structure comprising a motion-aware network and a target-aware network.
• Enabled end-to-end training without post-processing for streamlined implementation.

Main Results:

• DANet demonstrated superior performance compared to state-of-the-art trackers on event tracking datasets.
• The tracker achieved high accuracy and efficiency, running at over 80 FPS on a single V100 GPU.
• Simultaneous exploitation of motion and contour cues proved effective in identifying targets amidst distractors.

Conclusions:

• The proposed DANet offers a significant advancement in event-based visual object tracking.
• The distractor-aware approach enhances robustness in complex and dynamic environments.
• DANet provides a highly accurate and efficient solution for real-time object tracking using event cameras.