摘要
与可见光相比,红外光在部分复杂环境下仍能保持追踪目标的图像捕捉能力。因此,红外光图像可帮助计算机视觉跟踪算法提高目标跟踪的精度和鲁棒性。但在真实的目标追踪序列场景中,跟踪画面还存在眩光、形变和镜头抖动等干扰。为抵御此类干扰,提出了一种使用多尺度选择注意力的红外可见光融合目标跟踪算法(Selective Kernel Attention Fusion Tracking Network,SKANet)。该算法利用多尺度卷积核以及通道选择注意力,提取不同尺度大小的目标特征并将模型权重聚焦于质量较高的特征图、降低干扰带来的不利影响,从而提高模型的跟踪性能。通过在RGBT234和GTOT数据集上的验证结果表明,该算法可有效抵抗画面中干扰情况带来的不利影响,实现高精度的目标追踪。
Compared with visible images,thermal images have stable ability to catch the object under some complex situations.Therefore,introducing thermal image is important to improve tracking precision and robustness for tracking methods.However,there are various interferences of real-world scenarios(such as light occlusion,deformation and camera motion).To solve this problem,Selective Kernel Attention Fusion Tracking Network(SKANet)is proposed to achieve the RGB and thermal(RGBT)tracking.SKANet utilizes the multi-scale kernels and channel selection attention to extract the multi-scale features and reselect the robust feature based on channel dimension,which can effectively overcome the interferences and improve performance for RGBT tracking.Experimental results on RGBT234 and GTOT show that SKANet alleviate deteriorations introduced by the interferences and achieve a precision and robust performance for RGBT tracking.
作者
晏开祥
周冬明
王长城
周子为
YAN Kaixiang;ZHOU Dongming;WANG Changcheng;ZHOU Ziwei(School of Information science and Engineering,Yunnan University,Kunming 650500,China)
出处
《无线电工程》
北大核心
2023年第10期2261-2269,共9页
Radio Engineering
基金
国家自然科学基金(62066047,61966037)
云南大学研究生科研创新基金资助项目(KC-22222455,KC-22221913)。