Visual Tracking

Tracking objects in videos is a challenging and fundamental problem in computer vision. The objects can have changes in appearance and scale, and may be occluded. We are developing a number of tracking methods to tackle the main problems in visual tracking, such as handling contaminated observation, re-learning appearance model over time using deep learning techniques, and studying the sampling function in active learning frameworks. Our trackers outperform the state the art methods in the latest benchmark dataset.

Challenge:

• Using hand-crafted, rigid, non-adaptive feature representations is a limiting factor
• Erroneously included background samples in the object model
• Noisy observation sequences

 
Our Approach:

• Deep learning for automatically relearning feature representations and accurately adapting changes
• Advanced sampling functions using active learning framework
• Online robust PCA for updating of observation model

 

 
Research Outcomes:

• The state-of-the-art visual tracking methods that outperform all the rival approaches in the latest benchmark datasets
• Multiple conference and journal publications
• Invited keynote talk

 
Current Work:

• Developing the most accurate tracking algorithms for environment, safety, autonomous driving, and surveillance.
• Incorporating tracking into commercially valuable applications
• Continuing to play important role in visual tracking scientific community