Learning a Unified Tracking-and-Detection Framework with Distractor-aware Constraint for Visual Object Tracking

초록

Most of correlation filter-based trackers utilize the circulant structure of the training samples to learn discriminative filters to identify the tracked target, which have shown excellent performance in terms of both tracking accuracy and speed. However, CF-based trackers possess two potential drawbacks: the search region are limited to small local neighborhood for high-speed tracking purpose, thus they usually have very few context information and tend to drift from 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 paper, we propose a unified tracking-and-detection framework which involves both context learning and target re-identification with a target-aware detector to solve the above-mentioned drawbacks. We first incorporate the distractor constraint as context knowledge into a continuous correlation filter for distractor-aware filter learning. Then a SSD-based target-aware detector is trained by domain-specific meta-training (DSMT) approach for deep detection features and hard-negative samples generation. Moreover we propose the spatial-scale consistency verification method for target re-identification task. Compared with existing state-of-the-art trackers, UTDF-DA (ours) tracker can achieve improved tracking performance in terms of both accuracy and robustness, we demonstrate its effectiveness and efficiency with comprehensive experiments on OTB-2015, VOT-2016, and VOT-2017 benchmarks.