Learning a unified tracking‐and‐detection framework with distractor‐aware constraint for visual object tracking

Most of the correlation filter (CF)-based trackers utilise the circulant structure of the training samples to learn discriminative filters to identify the tracked target, which has shown excellent performance in terms of both tracking accuracy and speed. However, CF-based trackers possess two potential drawbacks: the search regions are limited to the small local neighbourhood for the high-speed tracking purpose; thus, they usually have very few context information and tend to drift from a target in extreme attributes, e.g. background clutter, large-scale variation, and fast motion. Another is that once the tracking target is lost under large displacement motion, it cannot be re-identified in subsequent frames. In this study, the authors propose a unified tracking-and-detection framework with distractor-aware (UTDF-DA), which involves both context learning and target re-identification with a target-aware detector to solve the drawbacks mentioned above. They first incorporate the distractor constraint as context knowledge into a continuous correlation filter for distractor-aware filter learning. Then a single-shot multibox detector-based target-aware detector is trained by domain-specific meta-training approach for deep detection features and hard-negative samples generation. Moreover, they propose the spatial-scale consistency verification method for the target re-identification task. Compared with existing state-of-the-art trackers, UTDF-DA (the authors’) tracker can achieve improved tracking performance in terms of both accuracy and robustness; they demonstrate its effectiveness and efficiency with comprehensive experiments on OTB-2015, VOT-2016, and VOT-2017 benchmarks .