A One-stage Infrared Dim Small Target Tracking Method by Fusing Spatiotemporal Information

The susceptibility of infrared dim small targets to background clutter during motion causes existing Siamese network tracking methods to fail to track the target consistently and stably. In light of this, this paper puts forward a means that integrates the spatio-temporal information of targets, and reconstructs the Siamese network tracking framework for the task of tracking infrared dim and small targets. The deficiencies of the original two-stage Siamese network trackers are remedied by this framework. By using deep learning networks, it simultaneously accomplishes feature extraction and target matching, attaining multi-layer feature maps. As the various layers within the multi-layer feature maps concentrate on diverse aspects, a channel and spatial attention approach is adopted to fuse these feature maps with respect to semantics and particulars. Finally, to further improve tracking accuracy, the multi-head attention mechanism is combined with the Convolutional Neural Network. In an anchor-free form, it directly predicts target classification and regression scores to complete the tracking task. The realization results show that the proposed infrared dim small target tracking method has better performance on the public infrared dim small target tracking dataset, it achieved a success rate with an Area Under the Curve of 58.7% and an accuracy rate reaching 78.6%.