Learning Adaptive Sparse Spatially-Regularized Correlation Filters for Visual Tracking

The correlation filter(CF)-based tracker is a classic and effective model in the field of visual tracking. For a long time, most CF-based trackers solved filters using only ridge regression equations with l2 -norm, which can make the trained model noisy and not sparse. As a result, we propose a model of adaptive sparse spatially-regularized correlation filters (AS2RCF). Aiming to suppress the noise mixed in the model, we improve it by introducing an l1 -norm spatial regularization term. This converts the original ridge regression equation into an Elastic Net regression, which allows the filter to have a certain sparsity while maintaining the stability of model optimization. The entire AS2RCF model is optimized using alternating direction method of multipliers(ADMM), and quantitative evaluations through extensive experiments on OTB-2015, TC128 and UAV123 demonstrate the tracker's effectiveness.

Introduction

Visual object tracking is a fundamental problem in the field of computer vision, requiring the tracker to keep tracking the target in subsequent frames after providing the target's initial state in the first frame, which is challenging because the target may be constantly deformed and disturbed by the complex environment during its motion. CF-based trackers [1], [2] are a type of tracker that updates online and usually sets the objective function as a ridge regression problem, using a circulant matrix structure to simplify the computation.