Object Tracking Algorithm Using Siamese Attention Mechanism based Key Point Network

Abstract

The tracker based on Siamese network usually adopts the cross-correlation of convolutional features between the object branch and the search branch to describe the visual object tracking task as a similarity matching. However, most trackers apply backbone networks to represent multi-scale features in a hierarchical manner, which has limitations in distinguishing similar object. In addition, the feature information on each channel and spatial region is treated equally in the calculation process. In fact, the feature importance of these locations is not the same, and the anchor box detection mechanism requires setting a large number of anchor boxes with redundant hyperparameters, which can lead to computational and memory overload. As for these problems, this paper proposes an object tracking algorithm using siamese attention mechanism based key point network, called SAMR (Siamese Attention Mechanism Reppoints). First of all, the backbone network res2net is used to represent fine-grained multi-scale features on a hierarchical basis for increasing the receptive field of each layer. Then the channel and spatial attention information of the attention mechanism is to enhance the discriminant ability and positioning ability of the feature map. Finally, a set of key points anchor-free frames is adopted to detect and track the object. Experiments on challenging datasets such as VOT2018, VOT2019, and OTB100 illustrate that the SAMR proposed in this paper is superior to other advanced trackers in tracking accuracy.