Accurate Visual Tracking Based on Discriminative Correlation Filter and Depthwise Cross-correlation

The combination of a discriminative correlation filter (DCF) and a Siamese network in correlation operations plays a pivotal role in obtaining multi-channel correlation information from two feature maps, i.e., a template and a search region. DCF trackers are popular for their robustness, but they are limited to bounding box tracking and have restricted transformation models. The Siamese network tracker, grounded in segmentation, significantly enhances tracking performance. We fuse DCF-based tracking with depthwise cross-correlation (DW-XCorr) to effectively utilize location and multi-channel semantic features. We integrate multi-channel semantic features from DW-XCorr with foreground and background features of the target, and input to a refinement module for producing pixel-wise target state predictions. We develop a weighted multiscale convolution block to expand the global receptive field within the refinement network. The efficacy of the proposed tracker is validated by experiments on eight visual tracking benchmarks, where we realize new state-of-the-art performance on the VOT2016 benchmark.