STASiamRPN: visual tracking based on spatiotemporal and attention

Visual tracking is an important research topic in the field of computer vision. The Siamese network tracker based on the region proposal network has achieved promising tracking results in terms of speed and accuracy. However, for fast-moving objects, the structure of the tracking system mainly focuses on information regarding the object appearance, ignoring information related to movement and change at any moment. The original 2D convolutional neural network cannot extract the spatiotemporal information of tracking object and cannot pay attention to the features of tracking object. In this research, a new tracking method is proposed that can extract the spatiotemporal features of tracking objects by constructing a 3D convolutional neural network and integrating the cascade attention mechanism and distinguish similar objects by background suppression and highlighting techniques. To verify the effectiveness of the proposed tracker (STASiamRPN), experiments on the OTB2015, GOT-10K and UAV123 benchmark datasets demonstrated that the proposed tracker was highly comparable to other state-of-the-art methods.