An Approach for RGB-T Object Tracking Based on Siamese Network
Xingya Yan, Xibo Ke · 2023
RGB images can capture rich object information but are susceptible to environmental influences. Thermal images are insensitive to illumination changes and have strong haze transmission capability, but lack detailed texture information of the object. RGB images and thermal images can be used in combination to provide complementary information for object tracking. To facilitate the information dissemination between the two modalities and suppress the background noise, we design the attention mechanism module to extract the common and individual features between infrared and visible image pairs, and propose an RGB-T bimodal fusion of the area proposal network. Experimental results show that our tracker achieves competitive results and a tracking speed of more than 120 frames per second.