Yang Zheng, Liu Kaiwen, Li Quanlong, Hou Yandong, Yan Zhiyu
School of Electrical Engineering, Yellow River Conservancy Technical Institute, Dongjing street, Kaifeng, 475004, Henan, China.
School of Artificial Intelligence, Henan University, Mingli street, Zhengzhou, 450000, Henan, China.
Heliyon. 2025 Jan 6;11(2):e41612. doi: 10.1016/j.heliyon.2024.e41612. eCollection 2025 Jan 30.
To enhance the correlation of feature information and enrich the pattern of cross-correlation metrics, we propose the Siam ST algorithm, which is based on a three-stage cascade (TSC) architecture. The sliding window is introduced in the last three layers of convolution blocks, which can obtain the global information of images and fully capture the target feature. The TSC structure is developed by using the regional proposal network. It makes the features of the current frame interact with the previous frame. As a result, our method has a high effect of robustness and association features extraction. Therefore, our ablation experiments are conducted on the VOT2016 dataset, and comparison experiments are conducted on four datasets, VOT2018, LaSOT, Tracking Net, and UAV123. Our proposed algorithm demonstrates a significant improvement compared to SiamRPN++ across four datasets.
为了增强特征信息的相关性并丰富互相关指标的模式，我们提出了基于三级级联（TSC）架构的暹罗时空（Siam ST）算法。在卷积块的最后三层中引入滑动窗口，其可以获取图像的全局信息并充分捕捉目标特征。TSC结构是通过使用区域提议网络开发的。它使当前帧的特征与前一帧进行交互。因此，我们的方法在鲁棒性和关联特征提取方面具有很高的效果。因此，我们在VOT2016数据集上进行了消融实验，并在四个数据集VOT2018、LaSOT、Tracking Net和UAV123上进行了对比实验。与SiamRPN++相比，我们提出的算法在四个数据集上均有显著改进。
