Thermal Infrared Object Tracking Based on Adaptive Feature Fusion

Abstract

SiamRPN++ has achieved excellent performance on thermal infrared object tracking. However, it directly fuses multi-layer features using weighted summation, which has the problem of insufficient feature fusion. In this paper, we propose an adaptive feature fusion module. It can fuse the features of different layers by adaptively allocating channel weights. Meanwhile, CIoU loss is used to make the regression of the bounding box more accurate. Experimental results show that the proposed method improves the baseline algorithm effectively and achieves excellent tracking accuracy and efficiency. The proposed method has strong robustness, effectively dealing with some challenges such as interference and occlusion. Therefore, the proposed method is valuable in practical application.