Abstract
<title>Abstract</title> <p> <italic> <bold>The visible-infrared target tracking has the problem of modal failure in the low light condition, occlusion and thermal interference. In complex scenarios, to improve the accuracy in tracking results, we built a dataset of 41, 000 synchronized video frames and proposed a drift warning method based on Siamese Transformer, crossmodal attention, temporal consistency constraint, Monte Carlo Dropout and temperature calibration. The experimental results indicate that the model has an average success rate of 81.2% in detecting the drifting, the precision of 86.5%, and the AUC of 0.887 in drift detection with an average warning trigger time of 8.4 frames before the drifting happens. The results show that this approach can dynamically evaluate the reliability of the modes and reduce the possibility of a wrong continuous tracking.</bold> </italic> </p>