Infrared Small Target Tracking Based on OSTrack Model

Abstract

Infrared small target tracking plays an important role in military reconnaissance, early warning, video surveillance, and civil applications. For tracking small infrared targets in this paper, a one-stream deep learning model is utilized. In order to integrate the processes of feature extraction and feature fusion, the model uses transformer as the framework’s major component and creates a bidirectional information flow between the template and the search picture pairs in the feature extraction stage. Use the head of the model to get the target position. Finally, post-processing of the target area, including tracking success, saves the coordinate information of the target frame; tracking failure, near, in the middle, and far from the target box, searches for the real target. It helps to solve the situation where the target moves fast and encounters a complex background to achieve better tracking results. It is tested on an infrared small target data set, and the results show that the method in this paper reaches 80.50% average tracking accuracy. The image sequences in the data set include sky, sea, and buildings. Tracking video and original images are shown at https://github.com/AHUT507LAB/Infrared-dim-small-target-tracking.

Keywords

• Infrared small target
• small target tracking
• post-processing
• deep learning