Visual Multi-Object Tracking through Deep Learning

Zusammenfassung

This thesis presents novel deep-learning approaches for tracking multiple simultaneous objects in videos, which is an essential component in several applications such as robotics or video surveillance. Traditional multi-object tracking systems, which rely on frame-by-frame detections and primarily geometric attributes, are ill-suited for real-time environments and open-set scenarios. To overcome these limitations, we introduce SiamMT, an innovative architecture that adapts single-object tracking techniques for handling multiple arbitrary targets in real time. This approach is further refined with SiamMOTION, which effectively manages distractors and accommodates objects of varying sizes by extracting semantically-richer features and proposing more accurate search areas. Lastly, this thesis proposes a Transformer-based architecture that complements a lightweight detector by recovering undetected objects for enhanced performance in multiple object tracking.