Multi-Object Tracking in Complex Scenes
Extending human perception through innovative tracking algorithms
"The real voyage of discovery consists not in seeking new landscapes, but in having new eyes." - Marcel Proust
Extending human perception through innovative tracking algorithms
"The real voyage of discovery consists not in seeking new landscapes, but in having new eyes." - Marcel Proust
Multi-Object Tracking (MOT) is a fundamental computer vision challenge that requires tracking multiple objects simultaneously throughout a video sequence. This is essential for applications like autonomous driving, surveillance systems, and sports analytics.
"The real voyage of discovery consists not in seeking new landscapes, but in having new eyes." - Marcel Proust
My research mission is to extend human perception through intelligent visual tracking systems — enabling machines to see, reason, and remain aware where human attention cannot.
DragonTrack: Transformer-Enhanced Graphical Multi-Person Tracking in Complex Scenarios
Published in WACV 2025
DragonTrack introduces a novel approach to multi-object tracking by combining transformer-based detection with graph modeling for inter-object relationships. Our framework achieves state-of-the-art performance on standard benchmarks while maintaining efficient computational requirements.
Key innovations:
MOTE: More Than Meets the Eye
Optical Flow-Based Multi-Object Tracking with Prolonged Occlusion Handling
Under Review, International Conference on Machine Learning (ICML), 2025
MOTE builds upon traditional tracking frameworks by incorporating optical flow and softmax splatting for disocclusion features. This approach significantly reduces identity switches in scenes with prolonged occlusions.
Key contributions:
UniTrack: A Differentiable Graph-Based Loss for Robust Multi-Object Tracking
Under Review, International Conference on Computer Vision (ICCV), 2025
UniTrack presents a unified tracking framework with a graph-based differentiable loss function that can be integrated with existing tracking architectures. This approach eliminates the need for scenario-specific tracking systems.
UniTrack delivers the following advantages:
LAPA is our current research focus for NeurIPS, introducing a novel attention mechanism for multi-object tracking that combines local and global context information to improve tracking accuracy.
Key innovations in LAPA include:
Current Status: Under development for NeurIPS submission
MOTE preserves tracking during occlusions using optical flow techniques
DragonTrack reduces identity switches in dense environments
UniTrack maintains stable tracking over extended video sequences
Plug-and-play optimization for MOTR, Trackformer, and FairMOT
Exploring attention mechanisms for ambiguous re-identification cases
Improving identity retention across extended temporal gaps
Integrating motion patterns to distinguish visually similar subjects
Robust identity association across multiple camera views with minimal calibration
Our research creates bridges between cutting-edge computer vision technology and life-changing applications that impact healthcare, safety, and human potential.
