Visual Tracking with Multiview Trajectory Prediction.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Visual tracking systems face challenges with occlusion and changing viewpoints.
• Existing multi-camera systems often require specific targets, static cameras, or calibration.

Purpose of the Study:

• To develop a versatile generic multiview tracking (GMT) framework.
• To enable robust tracking across multiple moving cameras without specific object models or calibration.

Main Methods:

• A cross-camera trajectory prediction network (TPN) dynamically encodes camera relations to handle occlusion.
• A collaborative correlation filter (CCF) is shared across cameras for robustness against view changes.
• Features are trained offline using single-view datasets within a correlation filter tracking framework.

Main Results:

• The proposed GMT framework demonstrates significant robustness improvements over state-of-the-art methods.
• Experiments were conducted on a newly created generic multiview tracking dataset (GMTD) and the CAMPUS dataset.
• The system effectively addresses occlusion and view changes in complex scenarios.

Conclusions:

• The developed GMT framework offers a generalized solution for multi-camera object tracking.
• The TPN and CCF components effectively mitigate common tracking challenges.
• This approach enhances the applicability of visual tracking in real-world, dynamic environments.