Deep Neural Networks for Visual Object Tracking: An Investigation of Performance Optimization

Abstract

This thesis advances visual object tracking by introducing four key enhancements that address current limitations in training data, network architecture, and tracking methodologies. It proposes a refined sampling strategy for Siamese Networks to enrich training data and develops a more efficient Partially Siamese Network through neural architecture search, achieving superior performance on benchmarks. The work further streamlines tracking with a new transformer-based pipeline and breaks ground with a speech-guided tracking framework, improving human-machine collaboration. These advancements are thoroughly validated, marking significant progress in the visual object tracking domain.