摘要
目前大多数热红外(TIR)目标跟踪算法都是基于相关滤波或者使用彩色跟踪器的模型进行特征提取。然而,两者都存在适用于彩色目标跟踪却对红外目标特征不敏感的缺陷,导致无法良好地应用到红外目标跟踪。为此,提出一种基于全局感知的孪生神经网络的红外目标跟踪器。将孪生神经网络的后三层特征进行融合优化,得到新的特征,同时加入了由空间转换网络和通道注意力组成的空间感知模块,以得到全局范围内的有效信息,通过引入自注意力机制,使算法更加专注于提取目标的判别信息,最后对结果进行响应融合得到最终的响应图。在PTB-TIR红外目标跟踪评估基准上的实验结果表明,本文算法能够适应多样的红外环境,同时能够保持良好的跟踪速度(20.2 frame/s),实现对红外目标有效且稳定的实时跟踪。
At present, most thermal infrared(TIR) tracking methods are based on correlation filters or RGB trackers for feature extraction. However, both of them are only suitable for RGB object tracking but not sensitive to the TIR object features, thus failing to be applied to the TIR object tracking. To this end, a TIR object tracker based on the global-aware siamese neural network was proposed in this paper. First, the features from the last three layers of the siamese neural network were fused to obtain new features. Second, the spatial-aware module composed of the spatial transformer network and channel attention was added to get the global effective information. Simultaneously, the self-attention mechanism was introduced to make the algorithm more focus on extracting the discriminant information of the objects. At last, the final response map was acquired by response fusion of the results. The experimental results on the TIR pedestrian tracking benchmark(PTB-TIR) show that the proposed algorithm can adapt to a variety of TIR environments while maintaining a high tracking speed(20.2 frame/s), achieving effective and stable real-time tracking of TIR objects.
作者
李畅
杨德东
宋鹏
郭畅
Li Chang;Yang Dedong;Song Peng;Guo Chang(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300130,China)
出处
《光学学报》
EI
CAS
CSCD
北大核心
2021年第6期166-176,共11页
Acta Optica Sinica
基金
河北省自然科学基金(F2017202009)
河北省创新能力提升计划(18961604H)。
关键词
机器视觉
红外图像
目标跟踪
孪生神经网络
注意力机制
machine vision
infrared image
target tracking
siamese neural network
attention mechanism
