Learning adaptive distractor-aware-suppression appearance model for visual tracking

Abstract

Some algorithms based on Siamese networks aim to improve the representation of target by combining background and target information, but they seldom consider adjusting the influence of background distractors on appearance modeling. In this paper, we propose an adaptive distractor suppression appearance model for robust visual tracking. Firstly, to fully utilize the valuable clues provided by the background, a new distractor model is specially designed to determine the weight of each distractor based on the similarity between the distractor and the target. This model adaptively fuses the distractors according to their weights, thereby focusing on distractors that are highly similar to the target. Secondly, a distractor model transformation strategy is constructed to rank the influence of the distractor model on appearance modeling, which mines the similarity relationship between the background distractor and target using regularized linear regression, effectively controlling the influence of the distractor model. Finally, we unify them into a learning adaptive distractor-aware-suppression appearance model for improving the discriminant ability of the appearance model, which selectively introduces the distractor model to suppress distractors according to the intensity of the distractor, achieving robust tracking in the presence of background interference. Experimental results on six benchmarks demonstrate that the proposed tracker achieves excellent performance in various challenging tracking tasks, particularly when facing background interference, where the tracking precision and success rate of our algorithm reach state-of-the-art levels.