Thermal infrared object tracking via unsupervised deep correlation filters

Abstract

In thermal infrared (TIR) object tracking, the scarcity of annotated TIR image datasets hampers the performance of deep learning-based trackers. Therefore, most trackers still only use handcrafted features. To tackle this problem, we propose an unsupervised deep correlation TIR object tracking framework. Motivated by the forward-backward tracking consistency of robust trackers, we explore a novel approach that learns a lightweight feature extraction network for TIR images in an unsupervised manner. We follow a Siamese correlation filter framework to construct the network and train it based on sample quality-sensitive consistency loss. To the best of our knowledge, this is the first unsupervised end-to-end trained feature extraction network for TIR object tracking. At the tracking stage, to improve tracking performance, we utilize continuous multichannel correlation filters to combine CNN features with handcrafted features. Extensive experiments indicate that the proposed method outperforms state-of-the-art trackers with more than 10% relative gain in EAO, clearly showing its effectiveness and robustness.