Anti-occlusion tracker for infrared ground targets based on improved trajectory prediction network

Abstract

Aiming at the problem of infrared ground target tracking drift or even failure caused by background occlusion, we propose an anti-occlusion tracker for infrared ground targets based on improved trajectory prediction network. First, the occlusion judgment criterion is proposed to accurately determine the occlusion of the infrared ground target. Then, based on the BiTraP trajectory prediction network, on the one hand, by introducing velocity information through the Siamese network structure and using one-way prediction, the SiamUniTraP trajectory prediction network is proposed, which improves the precision of trajectory prediction. On the other hand, by improving the training method and application method, it can predict the trajectory of ground target more accurately. For short-term occlusions, the SiamUniTraP network is employed to predict the bounding box for target occlusion based on temporal context information. For long-term occlusions, the search expansion strategy is proposed to handle the accumulation of prediction errors caused by the lack of real target information. Finally, the “second determination” criterion is proposed to achieve accurate recapture of the target and return to normal tracking. Infrared image sequence tests are conducted under occlusion conditions. Compared with the SuperDiMP tracker, the proposed tracker improves the success rate by 5.2% and the accuracy by 5.9% under the OPE evaluation index. This indicates that the proposed tracker has good robustness for infrared ground target tracking in occlusion situations, especially in full occlusion situations.