Infrared Object-Tracking Algorithm With Temporal Features Fusion Siamese Network
Jianfeng Song, Xu Hao, G L Ren, Qiguang Miao
Abstract
Infrared object tracking is an important subfield of computer vision that has a wide range of applications in harsh tracking environments. However, the absence of color information and blurred video frames make infrared object tracking more challenging than RGB tracking. To make full use of temporal information in infrared (IR) video and improve tracking accuracy, we propose a temporal features fusion mechanism that can fuse target features extracted by Siamese networks at different moments while tracking, thus integrating the temporal information to improve the accuracy. Weights for each target feature are calculated separately based on channel and spatial aspects and then aggregated. This enables element-wise feature fusion with fewer parameters and reduced computational cost. In addition, we propose an online feature update method, which collects target information online through a target-specific network and evaluates the results of each tracking to obtain high-quality tracking samples for temporal feature fusion. Our method exhibits state-of-the-art tracking performance compared to other existing tracking methods. In addition, we analyze the influence of several main factors on temporal fusion networks through an ablation study.