基于孪生区域候选网络的目标跟踪模型

为了解决被跟踪目标因尺度、形状变化导致的跟踪效果变差的问题,本文提出一种基于孪生区域候选网络的目标跟踪模型,对孪生区域候选网络(SiamRPN)优化,升级特征提取基准网络,采取多层特征融合模式,引入注意力机制模块增强位置特性和通道特性,并应用检测领域提出的GA-RPN替换原有的RPN(区域候选网络).OTB2015和VOT2018数据集的实验结果显示,本文模型对OTB2015数据集成功率为0.678,准确率为0.882,与SiamRPN相比分别提高了3.7%,6.2%;对VOT2018数据集检测帧率为31FPS,平均重叠期望为0.402,与SiamRPN相比提高了4.9%,测试结果表明本文模型具备较高的跟踪精度和较强的抗干扰性,满足实时性需求.

Target Tracking Model Based on Siamese Region Proposal Network

In order to solve the problem of poor tracking performance caused by changes in scale and shape of the tracked target,this paper proposes a new target tracking model based onSiamese Region Proposal Network(SiamRPN).In this paper,SiamRPN is improved by upgrading the feature extraction reference network,adopting a multi-layer feature fusion module,introducing an attention module to enhance the location characteristics and channel characteristics,and applying the GA-RPN that has been proposed in the detection field to replace the original RPN(Regional Proposal Network).The experimental results of the OTB2015 and VOT2018 datasets show that the model proposed in this paper has a success rate of 0.678 and a precision of 0.882 on the OTB2015 dataset,it is increased 3.7% and 6.2% respectively comparedwith the results of SiamRPN,and a detection rate of 31 FPS and an expected average overlap of 0.402 on the VOT2018 dataset,it is increased 4.9%comparedwith the results of SiamRPN.In addition,the interference tests of the scale and shape changes are carried out on the video collected indoors.The experimental results show that the model has higher tracking accuracy and stronger anti-interference performance,and meets the real-time requirements.