激光与光电子学进展, 2021, 58 (20): 2015002, 网络出版: 2021-10-14
结合缓冲区与三元组损失的孪生网络目标跟踪 下载: 606次
Siamese Network Target Tracking Based on Buffer and Triplet Loss
机器视觉 孪生网络 区域提议网络 缓冲区 三元组损失 machine vision Siamese network region proposal network buffer module triplet loss
摘要
针对SiamRPN(Siamese Region Proposal Network)在目标被短时遮挡以及外观剧烈变化的情况下存在定位不准确的问题,提出一种结合目标跟踪缓冲区与三元组损失的目标跟踪算法。该算法首先将原有的固定模板改为动态模板,提升复杂环境下相似度判别的可靠性;然后在模板缓冲区稀疏地缓存目标外观以应对跟踪过程中非语义样本的干扰,增强目标跟踪的鲁棒性;最后应用三元组损失以充分利用目标的正负样本特征,使跟踪更加具有判别能力。使用OTB100数据集进行实验,结果表明所提算法的成功率曲线下面积较SiamRPN提高了0.021,平均中心位置误差降低了25.56 pixel,平均重叠率提高了25.2%。
Abstract
Aiming at the problem of inaccurate positioning of the SiamRPN (Siamese Region Proposal Network) when the target is temporarily blocked and the appearance changes drastically, a target tracking algorithm combining target tracking buffer and triple loss is proposed. First, the original fixed template is changed into dynamic template to improve the reliability of similarity discrimination in complex environment. Then, the image of the target is sparsely cached in the template buffer to deal with the interference of non-semantic samples in the process of tracking and enhance the robustness of target tracking. Finally, the triplet loss is applied to make full use of the positive and negative sample characteristics of the target to make the tracking more discriminant. Experimental results with OTB100 dataset show that compared with SiamRPN, the area under the success curve of the improved algorithm increases by 0.021, the average center position error decreases by 25.56 pixel, and the average overlap rate increases by 25.2%.
郭嘉, 王鹏, 杨永侠, 李晓艳, 邸若海, 李雪. 结合缓冲区与三元组损失的孪生网络目标跟踪[J]. 激光与光电子学进展, 2021, 58(20): 2015002. Jia Guo, Peng Wang, Yongxia Yang, Xiaoyan Li, Ruohai Di, Xue Li. Siamese Network Target Tracking Based on Buffer and Triplet Loss[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2015002.