LAUSR

Visual tracking is a fundamental and highly useful component in various tasks of computer vision. Recently, end-to-end off-line training Siamese networks have demonstrated great success in visual tracking with high performance in terms of speed and accuracy. However, Siamese trackers usually employ visual features from the last simple convolutional layers to represent the targets while ignoring the fact that features from different layers characterize different representation capabilities of the targets, and hence this may degrade tracking performance in the presence of severe deformation and occlusion. In this paper, we present a novel hierarchical attentive Siamese (HASiam) network for high-performance visual tracking, which exploits different kinds of attention mechanisms to effectively fuse a series of attentional features from different layers. More specifically, we combine a deeper network with a shallow one to take full advantage of the features from different layers and apply spatial and channel-wise attentions on different layers to better capture visual attentions on multi-level semantic abstractions, which is helpful to enhance the discriminative capacity of the model. Furthermore, the top-layer feature maps have low resolution that may affect localization accuracy if each feature is treated independently. To address this issue, a non-local attention module is also adopted on the top layer to force the network to pay more attention to the structural dependency of features at all locations during off-line training. The proposed HASiam is trained off-line in an end-to-end manner and needs no online updating the network parameters during tracking. Extensive evaluations demonstrate that our HASiam has achieved favorable results with AUC scores of $$64.6\%$$64.6%, $$62.8\%$$62.8% and EAO scores of 0.227 while having a speed of 60 fps on the OTB2013, OTB100 and VOT2017 real-time experiments, respectively. Our tracker with high accuracy and real-time speed can be applied to numerous vision applications like visual surveillance systems, robotics and augmented reality.