Robust Visual Tracking via Hierarchical Convolutional Features-Based Sparse Learning

In recent years, convolutional features have significantly advanced the Discriminative Correlation Filter (DCF) based trackers. In contrast to hand-crafted ones, features extracted from a CNN retain high spatial resolution while preserving semantic information. The improvements come at the risk of reduction in speed and over-fitting caused by the insufficiency of training data for tracking. In this paper, a novel Hierarchical Convolutional Features and Sparse learning based Tracker (HCFST) is proposed. We effectively tackle the issues of computational bottlenecks and over-fitting in the DCF formulation via the multi-task sparse learning. First, most of the noisy and irrelevant feature maps are safely removed for robust appearance modeling. Redundant features rejection effectively mitigates the redundancy among features from hierarchical layers of CNNs. Then a sparser updating scheme is further presented for conditional model update. Extensive experiments are performed on various challenging sequences from OTB50 and OTB100 datasets. The proposed HCFST performs favorably against state-of-the-art methods.